When Anthropic unveiled Mythos and Project Glasswing, the reaction was immediate and polarized. Some dismissed it as fear-driven marketing, while others treated it as a credible shift in the threat landscape.

Like with many things, the truth is probably somewhere in the middle. I wanted to test that for myself, and since I recently got access to OpenAI’s Trusted Access for Cyber program, I decided to take it for a spin.

GPT-5.4 identified the bugs and helped assemble a working exploit chain, but it wasn’t a simple “build me an exploit” prompt. Guiding it required domain knowledge, iterative probing, and knowing which paths were actually exploitable.

On modern browsers like Safari, exploitation is less about finding bugs and more about finding bugs that still matter after multiple layers of defense.

The bug I’m going to talk about today sits in a more interesting category. The bug itself looked contained, and in many ways it was. It did not provide a path to RCE or a sandbox escape. What it did instead was cross a different boundary entirely: it broke the Same-Origin Policy.

If you visited a malicious page from any Apple device, it could read authenticated cross-origin data from other sites you use, including access tokens and other sensitive data, making account takeover trivial.

The video below shows the PoC we sent Apple, demonstrating leakage of sensitive data from both Apple Connect and iCloud / Apple ID endpoints. Although this demo focuses on Apple services, the issue affects all websites. This means that by visiting a malicious website, sensitive data from other domains is at risk of being leaked.

The Sandbox Russian Doll

Browser exploitation in 2026 is a lot like being trapped in a Russian doll.

You start in the smallest doll, and every time you escape one layer you discover you are still trapped inside another one.

Finding a low-level memory bug is not the same thing as finding an exploit. Most of these bugs die in the gap between “memory corruption happened” and “something meaningful crossed a security boundary.”

On the outside you have the browser process model. Even if renderer code goes wrong, the browser is trying very hard to keep that damage inside the web content process.

Inside that you have the web security model: Same-Origin Policy, CORS, opaque responses, cookie scoping, and credential modes. Even if a page can trigger a cross-origin request, the renderer, and especially the Gigacage, should not be able to access the response bytes. Right?…

The Bug

The original bug lives in the refresh logic for non-shared resizable WebAssembly memory.

When a non-shared WebAssembly.Memory grows in BoundsChecking mode, JavaScriptCore can replace the underlying memory handle. That part is not the bug. The bug is what happens after that to the JS-visible resizable buffer returned by memory.toResizableBuffer().

The bug is simple enough that once I saw it, it was hard to unsee it. Safari’s grow path effectively does this:

And the refresh step effectively does this:

After memory.grow(), WebKit updates the buffer metadata, but leaves m_data pointing at the old freed allocation.

So after a grow, JavaScript can hold a buffer whose reported size is new, whose handle is new, but whose actual data pointer still references the old freed Primitive Gigacage allocation.

That turns into a stale typed-array window over freed memory.

On its own, this is already a real bug. But we’re still stuck inside the JavaScriptCore gigacage, effectively sandboxed. Without a second bug to break out into the renderer, it doesn’t chain into anything meaningful. What we have is a solid first-stage primitive, but no real security impact on its own.

Why it did not look exploitable at first

The stale window is confined to the Primitive Gigacage, which immediately limits what you can do with it. Many typical targets either never land there, lack useful structure, or fail to produce any cross-boundary effect.

So early on, it had all the hallmarks of a bug that looks promising but rarely goes the distance:

• easy source-level root cause
• visible stale memory behavior
• real reclaim
• no clean escape path

This is where a lot of low-level browser bugs die.

What changed the problem was a very different framing: maybe I did not need to escape the cage at all.

Maybe I just needed the browser to place something valuable inside it.

The Pivot

Instead of asking “how do I get from my stale WASM view to some protected browser state?” I started asking a better question:

“What browser code takes data that JavaScript is not allowed to read, but still copies that data into normal renderer memory?”

Because that is all I need.

I don’t need to break the abstraction.

I just need the browser to break it for me.

That naturally narrows the search space to subsystems that:

• handle sensitive cross-origin data, and
• still allocate ArrayBuffer-backed memory as part of their internal pipeline

That points straight at Fetch. The Fetch API clearly indicates that the response is opaque, meaning that its headers and body are not available to JavaScript.

Opaque Responses Are Supposed to Be Opaque

At the API level, the Fetch model here is straightforward.

If I do a cross-origin request with:

fetch(url, { mode: “no-cors”, credentials: “include” });

The browser may send the request, including cookies depending on context, but JavaScript receives an opaque response.

That means:

• I can hold the Response object
• but I cannot read the body bytes

And WebKit enforces that in the obvious place:

FetchBodyOwner::readableStream() blocks opaque bodies via isBodyNullOrOpaque().

So at first glance, everything looks fine. The body is hidden. The policy is enforced. Same-Origin Policy survives another day.

Except it does not.

The Fetch Behavior that Broke the Modal

The surprising part is Response.clone().

If FetchResponse::clone() is called while the response is still loading, WebKit will internally create a readable stream so it can tee the body between the original response and the clone.

That internal path does not apply the same opaque-body check first.

And once that happens, hidden response bytes start becoming very real renderer objects.

This is the part that made me stop and stare at the source, because the mismatch is right there.

The normal body path blocks opaque responses:

But FetchResponse::clone() does this while the response is still loading:

That is why it works.

The visible accessor path says “opaque bodies do not get a stream.” The clone path says “if it is still loading, create a stream so both clones can tee it.”

That second path is exactly what I needed.

The data flows through normal ArrayBuffer creation paths:

• buffered chunks go through tryCreateArrayBuffer()
• later chunks go through takeAsArrayBuffer()
• shared buffer data gets copied into ordinary ArrayBuffer allocations inside the renderer

So the policy ends up split in two:

• the public Fetch API says the body is opaque
• the renderer still materializes the opaque body into readable byte arrays during clone-time streaming

Combined with the stale WASM window, it becomes a SOP break.

The Chain

At a high level, the exploit became:

1. Force the target WASM memory into the BoundsChecking path.
2. Call memory.toResizableBuffer().
3. Grow the memory.
4. Keep the stale resizable buffer whose pointer still targets freed Primitive Gigacage pages.
5. Trigger a cross-origin fetch(…, { mode: “no-cors”, credentials: “include” }).
6. Call response.clone() while the response is still loading.
7. Let Fetch internals materialize the hidden body bytes into ordinary renderer ArrayBuffers.
8. Reclaim the stale WASM-covered pages with those allocations.
9. Read the cross-origin bytes through the stale view.

That is the entire trick.

I never needed response.text(). I never needed response.arrayBuffer(). I never needed the public API to hand me the body.

The browser copied the body into memory for its own internal bookkeeping, and the stale WASM view read it directly.

That is why this bug stopped being “some weird WASM UAF” and became “this completely breaks the Same-Origin Policy.”

The file:// Detour

One of the weirdest parts of the research was that the request side behaved differently depending on where I launched it from.

In my testing, cross-origin requests were much easier to get moving from file:// than from a normal https attacker page.

That sounds backwards until you look at WebKit’s handling of local origins.

Document.cpp has explicit special-casing around local documents and settings like:

• allowUniversalAccessFromFileURLs
• allowFileAccessFromFileURLs

MiniBrowser exposes those knobs too, which made file:// very useful as a research environment. It let me focus on the memory side and confirm the leak path before I had a clean web-facing story.

But I did not want a local-file party trick.

I wanted a real web exploit.

And from a normal https page, the same request pattern was not giving me the reliability I wanted.

That is where about:blank saved me.

Why about:blank saved the final POC

The final PoC opens an about:blank popup and performs the fetches from there:

This ended up mattering a lot.

At first I thought this was just an origin-inheritance trick. That part is real:

So about:blank does inherit the opener’s origin.

But that alone does not explain why the popup path behaved differently.

What actually seems to matter is Safari’s cookie / first-party bookkeeping. Fetch subresource requests copy document->firstPartyForCookies() into the request:

And WebKit’s cookie blocking logic bails out immediately if that first-party domain is empty:

That is a very different path from a normal attacker-controlled https page. From a regular https://attacker.example origin, the first party is the attacker site, so a request to the victim site looks third-party and Safari’s tracking-prevention logic can suppress cookies.

From the about:blank popup path, the security origin still comes from the opener, but the popup’s top-level URL / first-party context is no longer a normal registrable https site in the same way. In practice, that was enough to make credentials: “include” requests behave differently and get me the authenticated traffic pattern I needed.

So the important point is not “about:blank disabled CORS.” It did not. The important point is:

• the popup kept the opener’s origin
• the request still went through normal Fetch/CORS code
• Safari’s first-party cookie logic treated that popup context differently

That was the difference between “cross-origin request happens but is useless” and “cross-origin request comes back with authenticated bytes worth stealing.”

Why this was fun

This is my favorite kind of browser bug.

Not because the root cause was complicated. It was not. The WASM bug was almost embarrassingly direct.

And not because the final chain was huge. It was not.

It was fun because it is exactly the kind of bug modern browser architecture is supposed to suppress.

A stale pointer inside a cage is supposed to stay a stale pointer inside a cage.

An opaque response is supposed to stay opaque.

Those are both reasonable assumptions.

The exploit works because both assumptions were true only locally.

JavaScriptCore gave me a stale view that looked hard to use. WebCore Fetch gave me sensitive bytes that looked impossible to read.

Put them together and Safari’s Same-Origin Policy fell apart.

Disclosure

We reported our findings to Apple. Shortly after, a fix shipped, suggesting the issue was already known internally.

The vulnerability (CVE-2026-20664) is addressed in iOS 26.4 and iPadOS 26.4 (23E6254 and later), and macOS Tahoe 26.4 (25E253 and later). Make sure your systems are up to date.

Closing Thoughts

The biggest thing on my mind after working with these models is the leverage they provide, and what that means for N-days. A security patch in popular software used to hide the underlying exploit behind time, effort, and expertise. Now that you can scale tokens instead of effort, that barrier is mostly gone.

This doesn’t turn exploitation into a trivial task. You still need someone who understands what they are looking at, can filter noise, and can steer the process when it stalls. But AI changes the unit of work. Instead of deep, sequential effort, you get parallel exploration and rapid iteration. The constraint shifts from raw effort to how effectively an operator can guide multiple lines of inquiry at once.
`

The post Hacking Safari with GPT 5.4  appeared first on Blog.

In today’s dynamic digital environment, the pressure to innovate has never been greater. Development teams are pushing for native cloud tools to maximize performance and cost-efficiency, while security teams require best-of-breed, enterprise-grade protection to defend against an ever-evolving threat landscape. This often creates a point of friction, forcing organizations into a difficult trade-off: sacrifice performance for security, or accept weaker protections for the sake of speed.

To resolve this challenge, Thales Imperva is collaborating with Google Cloud to deliver a solution that helps bridge this gap. We are proud to introduce Imperva for Google Cloud (IGC), an integrated security solution that offers the best of both worlds: enterprise-grade application security with the cloud-native performance you expect from Google Cloud.

Imperva for Google Cloud: A Holistic, Integrated Solution

Imperva for Google Cloud is not just another security layer; it is a fully managed, best-in-class Web Application and API Protection (WAAP) solution built directly into the fabric of Google Cloud. This integration, available now on Google Cloud Marketplace,   provides robust protection without disrupting your existing infrastructure or workflows.

• Cloud-Native Performance Without Compromise: Imperva for Google Cloud uses Google Cloud’s native Service Extension and Private Service Connect to inspect traffic within the Google Cloud network. This means all traffic analysis happens without your data ever leaving Google Cloud infrastructure, preserving optimal latency, performance, and data residency.
• Quick Deployment: Forget complex re-architecture. Imperva for Google Cloud can be deployed quickly using familiar tools like Terraform, Google Cloud CLI (gCloud CLI), or the Google Cloud console UI. There are no disruptive DNS, SSL, or network routing changes required, allowing you to achieve production-ready protection almost immediately.
• Enterprise-Grade Protection Out of the Box: Imperva for Google Cloud is powered by Imperva’s industry-leading security engine, delivering comprehensive WAF, advanced API Security, and Account Bot Protection. Backed by 24/7 threat research, the Imperva solution provides near-zero false positives, with 97% of customers successfully using default policies and 95% running in blocking mode from day one. This dramatically reduces the operational overhead of constant rule tuning.

Real-World Impact: Securely Accelerating Your Business

By eliminating the trade-offs between security and performance, Imperva for Google Cloud helps organizations achieve key business outcomes:

• Accelerate Lift-and-Shift Migrations: Migrate workloads to Google Cloud confidently with security that adapts to your applications, not the other way around. Eliminate migration delays caused by complex security re-architecture.
• Unleash DevOps-Friendly Security: Empower development teams to innovate at speed. IGC closes the security gaps in built-in tools without slowing down deployment velocity or requiring developers to become security experts.
• Protect Modern Cloud-Native Applications: Secure your Kubernetes and microservices architectures with best-in-class defenses optimized for low-latency environments.
• Achieve Unified Multi-Cloud Governance: Manage security for all your Imperva-protected environments from a single, unified dashboard, providing consistent policy management and visibility across your entire multi-cloud estate.

“Bringing Thales Imperva to Google Cloud Marketplace will help customers quickly deploy, manage, and grow the company’s integrated security solution on Google Cloud’s trusted, global infrastructure,” said Dai Vu, Managing Director, Marketplace & ISV GTM Programs at Google Cloud. “Thales can now securely scale and support organizations that want to use its Imperva for Google Cloud solution to increase protection for their cloud-native applications, APIs, microservices and more.”

×

May 21 Upcoming Webinar

Findings from the 2026 Bad Bot report

Register Now

Join Us on the Journey to More Seamless Cloud Security

As we approach key industry events like our exclusive Executive Briefing Center (EBC) meeting in late March and Google Cloud Next 2026 in April, the conversation around integrated  security has never been more relevant. The launch of Imperva for Google Cloud marks a pivotal moment in our relationship with Google, providing a clear path for customers to secure their digital assets without compromise.

Ready to secure your cloud-native applications?

• Request a demo: Experience IGC in action at Imperva for Google Cloud
• Start your evaluation: Available now on Google Cloud Marketplace

The post Enterprise-Grade Application Security, Cloud-Native Speed: Introducing Imperva for Google Cloud appeared first on Blog.

The recent buzz around Anthropic’s Mythos model has been intense, and for good reason. Early reports suggest a model that significantly advances automated reasoning over large codebases, vulnerability discovery, and exploit generation. Some are already calling it a “game changer” for offensive security. 

But like most breakthroughs in AI, the reality is more nuanced. 

Let’s unpack what Mythos is, why it’s getting so much attention, and where the real impact will (and won’t) be. 

What Is Mythos, and Why It Matters 

At its core, Mythos is designed to operate deeply within software systems: 

• It can reason across entire codebases, not just snippets  

• It demonstrates strong capabilities in multi-step vulnerability discovery  

• It can potentially chain findings into realistic exploit paths  

This is what sets it apart from earlier models. Traditional LLMs often struggled with: 

• Context fragmentation (limited memory of large systems)  

• Superficial pattern matching (vs. true reasoning)  

• Weakness in multi-stage attack logic  

Mythos appears to push beyond that, closer to what human security researchers do when analyzing complex systems. 

That’s the hype. Now let’s put it into perspective.

1. Closed Systems Still Have a Natural Advantage

One of the most important constraints, often overlooked, is access. 

Organizations running: 

• Licensed binaries  

• Closed-source products  

• SaaS platforms  

are inherently less exposed to this class of AI-driven analysis. 

Why? Because Mythos appears to be most effective when it has full visibility into the source code. Without that: 

• Reverse engineering binaries is still hard and lossy  

• SaaS environments expose only interfaces, not logic  

This creates a natural barrier for attackers. 

Although “security through obscurity” isn’t a solution, in practice: 

• Open-source projects and exposed codebases will feel the impact first  

• Closed vendors still need to worry, but they’re not suddenly transparent overnight 

2. The Real Pressure Point: Time-to-Mitigation

AI doesn’t just change what attackers can do, it changes how fast everything happens.  

And this is where security vendors feel the most pressure. The challenge isn’t whether vulnerabilities exist, it’s how fast vendors can respond once they’re discovered. 

The new race: 

• AI/ human finds vulnerability →  

• AI Exploit is generated quickly →  

• Attack traffic emerges earlier →  

• Defenses must adapt in near real-time.

This shifts the competitive advantage to vendors that can: 

• Automate security workflows to 

• Rapidly understand new attack patterns  

• Generate mitigations  

• Deploy protections before mass exploitation 

3. The Budget Reality: AI Red-Teaming Isn’t Cheap 

One of the least discussed aspects of Mythos is cost. 

Running such a model at scale involves: 

• High compute costs  

• Expensive infrastructure  

• For example, Anthropic admitted that “Across a thousand runs through our scaffold, the total cost was under $20,000” for finding vulnerabilities in OpenBSD.

• Significant human validation effort 

And that last part is critical. 

Every finding still requires: 

• Verification (is it real?)  

• Reproduction  

• Impact assessment  

Which means more security engineers per finding, not less.

Organizations will need to start budgeting for: 

• AI-assisted red teaming  

• Dedicated pipelines to process findings  

• Integration into SDLC workflows  

This mirrors what we’ve already seen with GitHub Copilot-style assistants and AI-based code analysis tools.

Implication for attackers: 

These “doomsday” capabilities are not evenly distributed. 

• Well-funded actors (nation-states, top-tier cybercrime groups) → likely adopters  

• Opportunistic attackers → much slower to benefit  

So the threat landscape widens at the top, not uniformly across all attackers.

4. Bug Bounty Programs Will Feel the Noise First

One immediate and very practical impact: bug bounty platforms are about to get noisy. 

Expect a surge of: 

• AI-generated vulnerability reports  

• Poorly validated findings  

• Duplicates and false positives  

This creates a scaling problem for security teams. 

Organizations will need to adapt: 

• Stronger triage filtering mechanisms (likely AI-driven)  

• Reputation systems for researchers  

• Penalties for repeated false positives  

• Potential adjustments in bounty pricing  

Otherwise, teams risk wasting cycles on low-quality reports and missing real vulnerabilities buried in noise. Ironically, AI will be needed to defend against AI-generated reports.

5. Not All Vulnerabilities Are Equal

Another important nuance:  

Finding a vulnerability ≠ exploiting it at scale. 

Even with Mythos: 

• Many findings will be low impact  

• Exploitation may require environment specific conditions  

• Real-world constraints (auth, rate limits, monitoring) still apply  

This is where traditional security layers still matter: 

• WAF, API protection, Bot protection 

• Identity protection 

• Data protection 

• Threat reputation 

Mythos increases discovery capability, but doesn’t eliminate defense in depth. 

Final Thoughts 

The Mythos model presents a meaningful step forward. It brings AI closer to acting like a real security researcher, capable of deep reasoning and complex analysis. 

But it’s not a universal “break everything” button. 

• Closed systems still provide friction  

• Costs limit widespread misuse  

• Defensive technologies remain highly relevant  

• Operational processes (triage, mitigation) become the real bottleneck  

The hype focuses on capability. The reality is about constraints and execution. 

And as always in cybersecurity, the winners won’t be those with the best tools, but those who can operationalize speed, from detection to mitigation, at scale. 

The post Anthropic Mythos: Separating Signal from Hype appeared first on Blog.

Executive Summary

In this article, we disclose a new high severity unauthenticated remote denial‑of‑service vulnerability we identified and reported in React Server Components that we’ve dubbed “React2DoS”.  In this blog, we’ll analyze its impact and place it in the broader context of recently found Flight protocol vulnerabilities, especially CVE‑2026‑23864.

Introduction

We are in a phase of the web where performance and developer experience are no longer trade-offs, they’re expectations. Modern frameworks compete to ship less JavaScript, reduce client-side complexity, and move logic back to the server.

React, as one of the dominant forces in frontend development, has been at the forefront of this evolution. With the introduction of React Server Components (RSC), the ecosystem embraced a new model: components that execute exclusively on the server, access databases and secrets directly, and stream a serialized UI representation to the client.

This architecture promises smaller bundles, cleaner separation of concerns, and more efficient rendering. Instead of hydrating everything on the client, Server Components emit a structured stream that the browser reconstructs locally.

At the heart of this mechanism lies a custom streaming protocol known as Flight. Through Flight, React can serialize complex structures, like arrays, maps, object references, even promises and async boundaries, allowing the server to describe rich UI trees in a compact format.

This is powerful.

But history has shown that when we introduce custom serialization formats and complex parsers, we also introduce risk. The server must deserialize and reconstruct object graphs from client-controlled input. And complex parsing logic has long been fertile ground for vulnerabilities.

In our research we discovered a denial-of-service vulnerability that allows an attacker to impose disproportionate computation to the remote server.

React2Shell and subsequent DoS vulnerabilities

Earlier this year, the disclosure of React2Shell caught much of the community off guard, triggering emergency patches and intense scrutiny of the React Server Components architecture, amplified by waves of low-quality AI-generated analysis that blurred the line between verified facts and speculation. This episode also prompted deeper investigations into and led to new discoveries related to the security of the Flight protocol and related parsing mechanisms.

CVE‑2026‑23864 (CVSS 3.1 of 7.5), stood out as a notable example and serves as a useful reference for understanding the mechanics behind the issue we explore in this research.

Among other vectors, this vulnerability concerned the BigInt deserialization path in Flight:

• $n markers denote BigInt values
• No limit was enforced on digit length

Therefore, sending a million‑digit BigInt could cause a significant computation cost, and CPU exhaustion. An example payload could look like this:

0:"$n9999999999...[repeated 1 million times]"

In our setup, a single query like this could delay the server’s execution by several seconds if the inbound payload reaches the maximum allowed size (1MB with Node.js runtime, 10MB with Edge runtime).

This was the starting point of our research, and we tried to find payload that would trigger a similar, or superior cost to the server. This is exactly what we found, actually more computationally-intensive  by several orders of magnitude.

React2DoS

React relies on a mechanism known as the React Flight Protocol to serialize values that are sent to Server Functions.

On the client side, data is transmitted to the server as small pieces (or “chunks”), for example through form submissions:

payload = {
  "0": (None, '["$1"]'),
  "1": (None, '{"category":"vehicle","model":"$2:modelName"}'),
  "2": (None, '{"modelName":"tesla"}'),
}

As illustrated above, these chunks can reference one another.

After deserialization on the server, the reconstructed object looks like this:

{ "category": "vehicle", "model": "tesla" }

At first, we tried to measure the cost of execution of every type of reference supported by the Flight protocol. Among them, we looked at two promising ones: $Q and $W, respectively instantiating new Maps and Sets from the client request payload.

The first observation we made was that it was possible to reference the root element in the root element itself (!), which paved the way to recursive expressions:

“0” : [“$Q0”]

This, would cause the execution of the following JavaScript expression:

New Map([null])

Which makes perfect sense, because at the time of resolution of $Q0, $0 is not known yet.

However, what surprised us, was the fact that the following expression:

“0” : [“$Q0”, “$Q0” ..., “$Q0”] (x n)

did trigger the execution of the Map constructor n times!

Indeed, the ReactFlightReplyServer uses a `consumed` attribute to prevent multiple computations of the same reference and prevent abuse. But this mechanism only enters in action when the reference is successfully resolved (see Fig 1).

Fig. 1: Exception doesn’t prevent recomputation of the same faulty Map 

Because the `new Map` expression failed (new Map([null]) is not a valid JavaScript expression), this outcome was not stored anywhere. But surprisingly, the deserialization is not interrupted by this exception! 

The execution of the expression `new Map ([null])` is pretty cheap, it takes our server around 0.03ms. Virtually instant. But this is neglecting the fact that  a threat actor can insert more than 100,000 instances in a 1MB payload, leading to the cost of several seconds, comparable to the CPU exhaustion issue behind CVE‑2026‑23864 and described above. 

Considering this, we submitted a first report to Meta, sharing this POC and demonstrating the impact. 

But soon after, we realized there was a way more impactful payload we could generate by exploiting our original idea.  

Instead of sending a series of “$Q0” that would immediately trigger the exception, we decided to introduce a series of valid map entries at the start of the root entry, to force the Map constructor to iterate over them before triggering the expected exception (see Fig. 2). 

Fig. 2: Internal recursive resolution of “$0” 

By doing so, we achieved a quadratic complexity, and a much more expensive payload ! The optimal number setting is n/2 valid maps and n/2 map references to the 0 object (“$Q0”). 

CVE‑2026‑23864 (CPU exhaustion) vs React2DoS (CVE-2026-23869) 

With our new attack vector, the computation could easily last several minutes. Therefore, with only small payloads of tens of kilobytes, it was possible to initiate impactful DoS attacks. 

To give ourselves an idea of the impact of this attack vector, we computed a chart showing the comparison between CVE‑2026‑23864 (CPU exhaustion) and React2DoS. The result showed that after only a few kilobytes, React2DoS starts to stand out, and when the payload size reaches hundreds of kilobytes, it is already more powerful by several orders of magnitude (see Fig. 3). 

Fig. 3: Comparison React2DoS – CVE‑2026‑23864 

Therefore, with a single request, a threat actor can trigger a computation that will take minutes to handle. By repeating this, complete denial of service can be achieved. 

Mitigation 

The React team fixed this issue via setting the consumed flag before any map/set constructor was called.

The issue affects React Server Components version 19.2.4 and below. We recommend that you update to the latest available version that patches this vulnerability as soon as possible.  

If your application already sits behind an Imperva proxy, it is automatically protected against this attack. 

Conclusion 

This case highlights an important reality: the path to innovation inevitably introduces complexity, and therefore risk. As ecosystems evolve rapidly, staying up to date and remaining aware of newly discovered security issues is essential. 

In a more personal way, it was a pleasure for me to delve into one of the most used framework in the world and discover a finding with meaningful impact. This wouldn’t have been possible if researchers before didn’t pave the way with their investigations and their recent findings (React2Shell,  CVE‑2026‑23864…).  

Disclosure Timeline 

Feb 3 2026 – Report including first payload 

Feb 5 2026 – Second payload reported 

April 8 2029 – Vulnerability fixed in 19.2.5 (patch backported to versions 19.0.5, 19.1.6)

The post React2DoS (CVE-2026-23869): When the Flight Protocol Crashes at Takeoff appeared first on Blog.

AI-driven automation is no longer emerging. It is already integrated and accepted as internet traffic. From AI assistants and crawlers to enterprise automation tools, websites are now routinely accessed by non-human actors operating at scale.  Vulnerabilities or weaknesses in your application infrastructure, including risky APIs, are no longer difficult to find, as agentic AI tools, paired with automation, can observe and test endpoints and access points faster than any human.

AI-aware bot protection is a security approach that detects, classifies, and controls automated traffic generated by AI agents, LLM-powered assistants, and autonomous tools — then applies granular policies based on each bot’s identity, intent, and behavior.

Key Takeaways:

• AI-powered bots now represent a significant and growing share of internet traffic, blending seamlessly into legitimate user sessions.
• Traditional bot detection cannot reliably distinguish between beneficial AI assistants and malicious AI-driven agents.
• Unmanaged AI bots create measurable business risks: analytics distortion, inventory manipulation, API abuse, account takeover, and content scraping.
• Imperva Advanced Bot Protection provides granular visibility and control over AI-driven traffic by tool type, category, behavior, and business function.
• Effective AI bot management in 2026 requires multilayered detection with real-time, policy-based response capabilities.

The challenge for security teams is no longer understanding why automation is increasing, but gaining clear visibility and control over what that automation is doing.

The result is a growing grey zone where distinguishing among human users, legitimate AI agents, and malicious bots becomes significantly more challenging, and where traditional security controls often lack the visibility needed to reliably distinguish among them.

According to Imperva’s 2025 Bad Bot Report, bad bots accounted for 32% of all internet traffic — a 2% increase year-over-year. With AI-powered tools accelerating automation, this figure is expected to grow significantly in 2026, making bot detection and bot management a critical priority for every organization.

How Do AI Bots Blend Into Legitimate Web Traffic?

AI agents and automated tools are improving how people interact with the internet, dramatically enhancing productivity and convenience. For example:

• AI assistants like ChatGPT, Perplexity AI, and Google Gemini retrieve real-time answers from multiple websites to summarise content or compare products
• Travel platforms continuously check flight prices, seat availability, and hotel inventory
• E-commerce monitoring tools track pricing, stock levels, and competitor offers across retailers
• AI-powered shopping assistants help users find deals or complete purchases faster
• Enterprise AI tools query SaaS platforms and APIs to automate workflows like reporting, customer support, and data enrichment
• Search and indexing bots extract and index web content to power AI-driven search experiences

However, the same technological advancements that enable these positive experiences are also empowering cybercriminals. Automation at scale lowers the barrier for malicious activity, putting malicious bots at a significant advantage when automated traffic is the expected baseline. They can blend seamlessly into legitimate traffic patterns, making detection significantly more challenging.

What Are the Business Risks of Unmanaged AI Bot Traffic?

Many organizations still view bot protection as optional. However, with AI agents such as crawler bots and fetch bots, now an accepted part of internet traffic and automation accelerating at scale, bot protection has become a core security requirement. Failing to treat it as such exposes organizations to serious business risks:

How Does Imperva Deliver AI Bot Detection and Control?

The ability to control which parts of your application functionality are accessible to AI tools is critical to your AI Security Strategy.

How Does Imperva Provide Visibility Into AI Bot Traffic?

Imperva Advanced Bot Protection (ABP) offers granular visibility into AI tools, agents, and crawlers, providing a detailed, real-time view of which AI tools are accessing your websites, applications, and API endpoints.

With ABP, security teams can clearly see which AI tools are hitting their environment, which applications and URLs are being accessed, the volume and frequency of requests, and whether those requests are being allowed, blocked, or challenged

This level of visibility ensures organizations know exactly what is interacting with their digital services and helps identify unintended policy outcomes, such as blocking AI tools they want to allow, or allowing tools they should restrict.

The AI Tools dashboard provides a centralized view of AI-driven traffic, enabling faster investigation and more informed decision-making.

How Can You Control AI Bots by Tool Type, Category, and Behavior?

Beyond visibility, Imperva enables precise control over how AI tools interact with your applications.

With ABP, security teams can easily:

• Allow, block, or rate-limit specific AI tools
• Apply policies based on categories such as AI crawlers, AI agents, and AI fetch bots
• Quickly adapt policies as new AI tools emerge

This allows organizations to move from reactive blocking to intentional control of automated access.

How Does Imperva Protect Critical Business Functions from AI Bots?

Imperva ABP also provides granular control at the application and business function levels, allowing organizations to define exactly which parts of their applications AI tools are allowed to access. This ensures that:

• Approved tools can only reach intended endpoints
• Sensitive paths, APIs, or business logic remain protected
• Access policies align with business and data governance requirements

This ensures AI tools interact with applications in a controlled, predictable, and secure way.

Why Is Imperva ABP a Leading Bot Management Solution?

ABP protection against AI builds on an already strong foundation of Advanced Bot Protection, combining multilayered detection, intelligent risk scoring, and real-time controls to accurately distinguish between human, legitimate automation, and malicious bots. With deep visibility, rapid decisioning, and expert support, ABP is already a proven solution for managing sophisticated bot threats. It is now further strengthened by the ability to monitor and control AI-driven traffic precisely.

Frequently Asked Questions About AI Bot Protection

Q: What is AI-aware bot protection?

A: AI-aware bot protection is a security approach that detects, classifies, and controls automated traffic from AI agents, LLM-powered assistants, and autonomous tools. Unlike traditional bot detection that relies on static signatures, AI-aware protection uses behavioral analysis and AI tool fingerprinting to distinguish between beneficial AI assistants, legitimate automation, and malicious bots.

Q: What is the difference between traditional bot detection and AI-aware bot management?

A: Traditional bot detection identifies bots using predefined signatures and rules, treating most automated traffic as either good or bad. AI-aware bot management goes further by classifying AI tools by type, category, and behavior — enabling organizations to allow helpful AI agents while blocking or rate-limiting harmful ones with granular policies.

Q: How do AI agents bypass conventional bot defenses?

A: AI agents can mimic human browsing behavior, rotate IP addresses, solve CAPTCHA, and generate realistic session patterns. Because they operate as legitimate AI tools (such as AI assistants and search crawlers), they often pass through conventional defenses that only look for known malicious signatures.

Q: What business risks do AI bots create?

A: Unmanaged AI bots can distort marketing analytics, hoard inventory, abuse API business logic, perform credential stuffing for account takeover, scrape proprietary data and competitive intelligence, and degrade customer experience through increased site latency.

Q: Can businesses allow some AI bots while blocking others?

A: Yes. Solutions like Imperva Advanced Bot Protection enable granular control, allowing organizations to allow specific AI tools (such as approved search crawlers), rate-limit others (such as AI assistants accessing content), and block malicious AI agents — all at the individual tool, category, or endpoint level.

Q: What is agentic AI, and why does it matter for application security?

A: Agentic AI refers to autonomous AI systems that can independently browse the web, interact with APIs, and complete multi-step tasks without human oversight. These agents can probe for vulnerabilities, test endpoints, and access business functions faster than any human, making agentic AI security a critical concern for organizations.

Monitor, Control, and Prevent AI-Driven Bot Threats

Automation is now a permanent and growing part of how the internet operates. The critical challenge is no longer detecting bots alone but understanding and controlling AI-driven interactions at scale.

Organizations need to know exactly which AI tools are accessing their environments, what they are doing, and how to control that access with precision.

Imperva Advanced Bot Protection delivers the visibility, control, and adaptive protection required to operate securely in this new environment.

By enabling organizations to monitor AI agents, control their access at a granular level, and prevent malicious automation from hiding within legitimate traffic, Imperva helps businesses confidently embrace the future of AI-driven digital experiences.

Learn how Imperva Advanced Bot Protection delivers AI-aware bot management for your applications. Explore our bot protection solutions or download the latest Imperva Bad Bot Report for the most current data on AI-driven bot threats.

The post Why AI Bot Protection and Control Are Essential for Application Security appeared first on Blog.

For years, a lot of risky APIs survived simply because they were hard to find. They weren’t documented. Only a handful of engineers knew the endpoints. And if an attacker wanted to abuse them, they had to spend real time reverse‑engineering traffic and guessing how things worked.

That “security by obscurity” was never a security strategy, but it did create friction.

AI removes that friction.

Today, coding assistants and agentic tools can observe patterns in traffic, infer undocumented endpoints, generate proof‑of‑concept exploits, and test thousands of permutations faster than any human. We’ve already seen what happens when exposed APIs meet automation at scale: a hobbyist was able to gain control of thousands of robot vacuums due to exposed APIs and an over‑privileged token, something that simply wouldn’t have scaled without automation on the attacker side.

The takeaway is straightforward: if you don’t know where your APIs are, what they expose, and who can talk to them, AI will find those gaps for you, either in the hands of your developers or your attackers.

Why has API security become critical in the age of AI agents?

API security is the foundation of protecting applications against automated, AI-driven threats. In the past, attackers relied on manual reverse-engineering to discover undocumented API endpoints. Today, AI agents and coding assistants can autonomously map traffic patterns, infer hidden endpoints, and test thousands of exploit permutations in seconds. Furthermore, AI agents can bypass traditional web application firewalls (WAFs) by executing perfectly formatted, syntactically correct requests that abuse business logic—such as chaining legitimate calls to perform a Broken Object Level Authorization (BOLA) attack.

Because AI agents use APIs as their primary control plane, securing these interfaces is no longer just about preventing data breaches; it is about establishing the necessary guardrails to ensure AI tools operate safely and within their intended scope.

How AI Agents Change the Threat Model

AI doesn’t just make attackers faster. It changes what “attack” looks like, because agents can behave like normal users while still doing abnormal things.

1) Business Logic is the New Frontline

Traditional API protections – gateways, WAFs, basic input validation, are good at stopping obviously bad traffic: missing tokens, malformed payloads, suspicious content types.

But agents don’t have to look suspicious. They can follow every syntactic rule and still abuse your business logic.

Imagine an agent that:

• Uses a valid user token and calmly walks the edges of a pricing API until it discovers discount combinations you never intended to allow.
• Chains perfectly legitimate calls to pivot from one customer data to another customer’s data. This effectively executes a Broken Object Level Authorization (BOLA) attack – a critical vulnerability highlighted in the OWASP API Security Top 10 – without brute‑forcing raw IDs.

Nothing in those requests’ screams “attack.” The danger is in the sequence, the intent, and the scale, the exact things many baseline controls don’t reason about.

2) Agent-Specific Protocols Expand the Attack Surface

Agents aren’t only calling the same APIs as your mobile app calls. They’re increasingly using agent‑first toolchains and protocols that wrap platforms behind “tool” interfaces, making discovery and invocation easier than ever.

Look at what’s happening across major SaaS ecosystems: new CLIs and frameworks are designed so an agent can discover capabilities, understand schemas, and call dozens of APIs through a single control surface. Under the hood it’s still JSON over HTTP but packaged in protocols and workflows many security tools don’t meaningfully parse or recognize.

If your security stack doesn’t understand what it’s looking at, it can’t apply real policy. It just sees “some JSON” and hopes for the best.

The Thales Vision: API Security as the AI Agents’ Control Plane

At Thales, we see API Security evolving into the control plane for AI agents: the place where you get a coherent view of what agents are doing, which APIs they’re touching, and how to govern that behavior, consistently and at scale.

1) Start with ruthless visibility

You can’t protect what you can’t see, and AI moves too fast for spreadsheets and tribal knowledge.

We’re focused on:

• Finding every API: Discovering shadow, zombie, and newly created APIs across clouds and data centers, then mapping the data they expose and the business functions they support.
• Making agent traffic visible: Identifying traffic that comes from agents and agent toolchains, tying it back to the human or system they’re acting for, and surfacing suspicious patterns early.

The goal: when your CISO asks, “Which agents can touch customer PII today?” you can answer with confidence instead of guesswork.

2) Speak the same language as AI agents

We’re extending the API Security engine, so it doesn’t just see “JSON over HTTP ” but understands the agent protocols layered on top, things like MCP (Model Context Protocol) style streams and backend API calls from an agent-oriented CLI.

Once we can parse and normalize that traffic, we can:

• Apply the same validation and anomaly detection we already use for REST and GraphQL.
• Correlate what an agent is doing across back‑end services, rather than treating every request as an isolated event.

In practice, that means the security brain becomes protocol‑aware. Whether an action comes from a mobile app, a browser, or an AI agent using a modern toolchain, the same set of eyes is watching.

3) Put real guardrails around tokens and delegation

Agents run on delegation. They act on behalf of users and services using tokens, keys, and temporary credentials. When those credentials are over‑privileged or long‑lived, you get “quiet catastrophe” scenarios, like a single token shared among thousands of agents.

We’re building on our existing token visibility to:

• Score token risk: Evaluate scope, lifetime, usage patterns, and anomalies like sudden geography changes or volume spikes.
• Create policies specifically for agent delegation: For example, “This support agent’s token can only read billing data for the current customer, up to N requests per hour, and never export full datasets.”
• Catch replay and abuse: Detect when tokens are cloned, reused from odd locations, or used by unexpected agent identities.

If an AI agent starts stretching beyond the intent of its access, querying too broadly, too often, or in the wrong context, the platform should be able to flag, throttle, or cut it off in real time.

4) Defend the messy middle: business logic and BOLA

Agents follow natural‑language prompts, not carefully designed UI flows. That makes them unusually good at stumbling into the “negative space” of your application: edge paths nobody documented, but your back end still accepts.

Our approach anchors security in behavior and intent:

• Model sequences of calls as workflows and look for patterns that don’t match real user behavior, for example, moving from one customer account to another without a corresponding permission to change.
• Treat BOLA as more than “did you increment an ID,” and start reasoning about what resource the agent is effectively asking for when it requests “all internal reports” or “all projects in the system.”

The endgame is business‑level guardrails you can express clearly, and enforce across all agents, regardless of how clever the prompts are.

Meeting you where you already are

None of this works if it requires an exotic, parallel deployment just for AI. That’s why we’re embedding agent controls into the places customers already rely on Imperva today:

• Imperva Cloud WAF for internet-facing API
• Imperva WAF Gateway for on-prem and hybrid environment
• Imperva eWAF for cloud-native and microservices workloads

In each case, it’s the same security engine doing heavy lifting, discovering APIs, understanding protocols, analyzing behavior, and enforcing policy inline on every agent’s call.

Where we’re heading

AI agents are already inside organizations, helping engineers, answering customers, and automating operations. The real question is whether they’re operating inside guardrails you actually understand.

Our view is simple:

• You don’t secure AI by bolting something onto the model.
• You secure AI by controlling the APIs and data the model can reach.

By turning API Security into the shared control plane for AI agents, across discovery, protocol understanding, token governance, and business‑logic protection, we want to help teams say “yes” to AI without crossing their fingers behind their back.

If you can see every agent, every call, and every token, you can turn AI from a wild card into an engineered advantage. That’s the future we’re building toward.

The post API Security for AI Agents: Why Protection Has Never Been More Important. appeared first on Blog.

Application development has changed dramatically. Enterprises now release software faster, operate more digital services, and deploy applications across a mix of public cloud, private cloud, APIs, containers, and on-premises infrastructure.

As application delivery has accelerated and architectures have become more distributed, a disconnect has emerged between the teams building applications and those responsible for protecting them.

This tension is often described as the Wall of Confusion between DevOps and IT Security.

But the challenge does not stop there.

Over time, organizations have also introduced multiple security tools to protect different parts of the application stack. Each tool is managed separately, often by different teams, through different platforms, policies, and workflows.

The result is an additional layer of complexity. Security teams must navigate multiple vendors and fragmented controls, while DevOps teams experience delays as security becomes harder to integrate into fast-moving development cycles.

Understanding how to break down both the organizational and operational layers of this confusion is essential for organizations that want to maintain innovation while ensuring consistent, scalable security.

Applications Now Run Across Hybrid Environments

Today, around forty percent of enterprise applications run in the public cloud, and that number is expected to rise significantly to 62% over the next two years.

Source: Vanson Bourne Survey, “DevOps vs Security: Breaking Down the Wall of Confusion in Modern Application Delivery”

Yet the shift to cloud does not mean applications live in one place. Most organizations now operate across hybrid and multi-cloud environments where applications run across public cloud platforms, private cloud infrastructure, on-premises systems, Kubernetes clusters, and an expanding network of APIs.

Cloud-agnostic strategies are also becoming more common as organizations seek flexibility and avoid dependence on a single provider. At the same time, many enterprises continue to operate legacy systems alongside modern cloud-native services.

The result is a highly distributed application landscape. Applications now run across multiple environments simultaneously, and security must be able to protect them wherever they operate.

Source: Vanson Bourne Survey, “DevOps vs Security: Breaking Down the Wall of Confusion in Modern Application Delivery”

DevOps and Security Want the Same Outcome

Despite the perception of conflict, DevOps and IT Security teams are largely aligned on the goals of modern application security. Both groups ultimately want the same outcome: applications that are secure, reliable, and able to scale with business demand.

Research conducted with Vanson Bourne reinforces this alignment. 96% of DevOps and 95% of IT Security professionals agree that modern environments require security that is flexible across any architecture.

This global study of 1,500 professionals across the US, Europe, and APAC highlights an important point. Modern application security is not just a technology problem. It is a workflow and collaboration challenge.

Security and DevOps want the same outcome, but they experience different frustrations. These gaps can create delays, bottlenecks, false positives, and friction that undermine the cloud-native innovation organizations are working to achieve.

The Wall of Confusion: Conflicting Priorities, Fragmented Security and Tool Sprawl

The Wall of Confusion is not just about DevOps and Security working in silos. It is also about how security is delivered. Over time, organizations have added more and more security tools. One for web applications, another for APIs, another for cloud, another for containers. Each tool solves a specific problem, but together they create complexity instead of clarity.

Security teams are left navigating multiple vendors, switching between management platforms, and maintaining different policies across environments. This makes it difficult to keep controls aligned and increases operational overhead.

At the same time, gaps begin to appear. As applications move across environments, it is not always clear if they are fully protected. Policies become inconsistent because what is set in one environment does not automatically apply to another.

In fact, based on a 2026 survey of Imperva Application Security customers, 77% of security professionals say operational complexity is their biggest challenge.

For DevOps teams, this complexity shows up as delay. Security becomes a bottleneck not because it is unnecessary, but because it is too difficult to operationalize.

That is the wall and it is what needs to come down.

Why Traditional Security Models Fall Short

When applications operate across multiple environments, security approaches designed for fixed infrastructure quickly become difficult to manage.

Many organizations rely on a mixture of embedded protections, centralized security services, and environment-specific tools to protect different parts of their application landscape. While each solution may address a particular need, together they can create fragmented security architectures. This fragmentation leads to inconsistent policies, duplicated alerts, limited visibility, and increased manual effort.

Security teams must manage multiple tools and workflows, while development teams experience delays when security is applied inconsistently or too late in the process. Both teams are constrained by the same underlying issue: security models that were not designed for modern, distributed application environments.

Security Must Move with the Application

Modern applications are no longer tied to a single infrastructure model. They are composed of microservices and APIs, deployed through automated pipelines, and distributed across multiple environments.

Security therefore cannot remain a centralized checkpoint that appears late in the development process. Instead, protection needs to move with the application and operate consistently wherever that application runs.

This means security controls must function across public cloud environments, private infrastructure, hybrid deployments, Kubernetes clusters, APIs, and the traditional systems that many organizations still rely on.

DevOps and IT Security teams increasingly recognize this shift. They are not asking for less security. They are asking for security that works the way modern applications work.

Securing Applications Anywhere with Thales

As application architectures continue to evolve, organizations are no longer dealing with a single security challenge, but with the need to protect applications consistently across every environment they operate in.

The issue is not just distribution. It is how to secure that distribution without adding more tools, more complexity, or more operational overhead.

Security strategies built around isolated environments or disconnected tools are no longer sufficient. What is needed is a unified approach that delivers consistent protection, visibility, and control across the entire application landscape.

Now, the question becomes how to deliver that in practice.

Many vendors talk about flexibility but still require organizations to choose a single deployment model or manage multiple disconnected solutions. Imperva takes a fundamentally different approach. It meets organizations where they are, supporting multiple deployment models while maintaining a single, unified security experience.

This includes protection for internet-facing applications and APIs through Imperva Cloud, native integration for public cloud environments (Imperva for Google Cloud), container-based deployment for Kubernetes and microservices, and gateway deployment for on-premises, hybrid, and air-gapped environments.

The key is that all of these deployment options are powered by the same Imperva Security Engine.

This means one management console, consistent policies across every environment, and unified visibility across the entire application portfolio, regardless of where applications are deployed. Security teams do not need to manage multiple tools or vendors, and DevOps teams do not need to change how they build and deploy applications.

That is what securing applications anywhere really means.

Download the whitepaper: DevOps vs Security: Breaking Down the Wall of Confusion in Modern Application Delivery

The post Securing Applications Anywhere: Breaking Down the Wall of Confusion appeared first on Blog.