The zero-day was designed to bypass 2FA and it was developed by a prominent cybercrime group.
The post Google Detects First AI-Generated Zero-Day Exploit appeared first on SecurityWeek.
The company topped revenue and earnings forecasts for the first quarter of 2026, but its shares plunged more than 20%.
The post Cloudflare Lays Off 1,100 Employees in AI-Driven Restructuring appeared first on SecurityWeek.
Last week on Malwarebytes Labs:
Stay safe!
Browse like no one’s watching.
Malwarebytes Privacy VPN encrypts your connection and never logs what you do, so the next story you read doesn’t have to feel personal. Try it free →
Researchers tracked a large AI‑themed investment scam campaign involving more than 15,000 domains. It uses cloaking and deepfakes to hide from security tools while targeting ordinary users.
Criminals abused the Keitaro ad-tracking platform as part of a cloaking system so real victims see scam content, while security scanners, ad reviewers, and some random visitors see harmless pages, making the operation hard to detect and shut down.
Keitaro is a commercial tracking platform originally meant for digital marketers to manage ad campaigns, test which ads work best, and route visitors to different landing pages.
Because it is feature rich, easy to spin up on regular hosting, and built to filter and route traffic, criminals found they can abuse those capabilities to run scams at scale.
Traffic starts in many places. The scammers used compromised websites, spam emails, social media posts, and online ads, all quietly routing through the same tracking infrastructure.
The scam sites typically promise “Smart AI Trading Technology” or “Intelligent Trading Solutions” and claim consistently high returns, often reinforced with deepfake images or fabricated media to look more credible.
Some parts of the campaign now use deepfake videos and fake interviews with well-known public figures, making it look like a celebrity, or finance expert personally endorses the platform.
Once you follow a link, the cloaking part of the operation kicks in. Cloaking is the trick that makes these scams so hard to see from the outside.
When you click an ad or link, your visit passes through a traffic distribution system (TDS), a kind of router for web visitors that decides which page you see. In these cases, the TDS is connected to the tracker.
The system checks things like:
You’re shown the real investment scam landing page only if you match the “ideal victim” profile (for example, a regular consumer in a target country coming from a social media ad).
Everyone else, like a security researcher, ad platform reviewer, or automated scanner, gets shown a benign page, like a generic blog or placeholder site.
The best way to stay safe is to stay informed about the tricks scammers use. Learn to spot the red flags that almost always give away scams and phishing emails, and remember:
Pro tip: Malwarebytes Scam Guard can help you recognize and analyze scams.
Stop threats before they can do any harm.
Malwarebytes Browser Guard blocks phishing pages and malicious sites automatically. Free, one click to install. Add it to your browser →
During Q1 2026, the exploit kits leveraged by threat actors to target user systems expanded once again, incorporating new exploits for the Microsoft Office platform, as well as Windows and Linux operating systems.
In this report, we dive into the statistics on published vulnerabilities and exploits, as well as the known vulnerabilities leveraged by popular C2 frameworks throughout Q1 2026.
This section provides statistical data on registered vulnerabilities. The data is sourced from cve.org.
We examine the number of registered CVEs for each month starting from January 2022. The total volume of vulnerabilities continues rising and, according to current reports, the use of AI agents for discovering security issues is expected to further reinforce this upward trend.
Total published vulnerabilities per month from 2022 through 2026 (download)
Next, we analyze the number of new critical vulnerabilities (CVSS > 8.9) over the same period.
Total critical vulnerabilities published per month from 2022 through 2026 (download)
The graph indicates that while the volume of critical vulnerabilities slightly decreased compared to previous years, an upward trend remained clearly visible. At present, we attribute this to the fact that the end of last year was marked by the disclosure of several severe vulnerabilities in web frameworks. The current growth is driven by high-profile issues like React2Shell, the release of exploit frameworks for mobile platforms, and the uncovering of secondary vulnerabilities during the remediation of previously discovered ones. We will be able to test this hypothesis in the next quarter; if correct, the second quarter will show a significant decline, similar to the pattern observed in the previous year.
This section presents statistics on vulnerability exploitation for Q1 2026. The data draws on open sources and our telemetry.
In Q1 2026, threat actor toolsets were updated with exploits for new, recently registered vulnerabilities. However, we first examine the list of veteran vulnerabilities that consistently account for the largest share of detections:
Among the newcomers, we have observed exploits targeting the Microsoft Office platform and Windows OS components. Notably, these new vulnerabilities exploit logic flaws arising from the interaction between multiple systems, making them technically difficult to isolate within a specific file or library. A list of these vulnerabilities is provided below:
These three vulnerabilities were utilized together in a single chain during attacks on Windows-based user systems. While this combination is noteworthy, we believe the widespread use of the entire chain as a unified exploit will likely decline due to its instability. We anticipate that these vulnerabilities will eventually be applied individually as initial entry vectors in phishing campaigns.
Below is the trend of exploit detections on user Windows systems starting from Q1 2025.
Dynamics of the number of Windows users encountering exploits, Q1 2025 – Q1 2026. The number of users who encountered exploits in Q1 2025 is taken as 100% (download)
The vulnerabilities listed here can be leveraged to gain initial access to a vulnerable system and for privilege escalation. This underscores the critical importance of timely software updates.
On Linux devices, exploits for the following vulnerabilities were detected most frequently:
Dynamics of the number of Linux users encountering exploits, Q1 2025 – Q1 2026. The number of users who encountered exploits in Q1 2025 is taken as 100% (download)
In the first quarter of 2026, we observed a decrease in the number of detected exploits; however, the detection rates are on the rise relative to the same period last year. For the Linux operating system, the installation of security patches remains critical.
The distribution of published exploits by software type in Q1 2026 features an updated set of categories; once again, we see exploits targeting operating systems and Microsoft Office suites.
Distribution of published exploits by platform, Q1 2026 (download)
We analyzed which vulnerabilities were utilized in APT attacks during Q1 2026. The ranking provided below includes data based on our telemetry, research, and open sources.
TOP 10 vulnerabilities exploited in APT attacks, Q1 2026 (download)
In Q1 2026, threat actors continued to utilize high-profile vulnerabilities registered in the previous year for APT attacks. The hypothesis we previously proposed has been confirmed: security flaws affecting web applications remain heavily exploited in real-world attacks. However, we are also observing a partial refresh of attacker toolsets. Specifically, during the first quarter of the year, APT campaigns leveraged recently discovered vulnerabilities in Microsoft Office products, edge networking device software, and remote access management systems. Although the most recent vulnerabilities are being exploited most heavily, their general characteristics continue to reinforce established trends regarding the categories of vulnerable software. Consequently, we strongly recommend applying the security patches provided by vendors.
In this section, we examine the most popular C2 frameworks used by threat actors and analyze the vulnerabilities targeted by the exploits that interacted with C2 agents in APT attacks.
The chart below shows the frequency of known C2 framework usage in attacks against users during Q1 2026, according to open sources.
TOP 10 C2 frameworks used by APTs to compromise user systems, Q1 2026 (download)
Metasploit has returned to the top of the list of the most common C2 frameworks, displacing Sliver, which now shares the second position with Havoc. These are followed by Covenant and Mythic, the latter of which previously saw greater popularity. After studying open sources and analyzing samples of malicious C2 agents that contained exploits, we determined that the following vulnerabilities were utilized in APT attacks involving the C2 frameworks mentioned above:
The nature of the described vulnerabilities indicates that they were exploited to gain initial access to the system. Notably, the majority of these security issues are targeted to bypass authentication mechanisms. This is likely due to the fact that C2 agents are being detected effectively, prompting threat actors to reduce the probability of discovery by utilizing bypass exploits.
This section highlights the most significant vulnerabilities published in Q1 2026 that have publicly available descriptions.
At the core of this vulnerability is a Type Confusion flaw. By attempting to access a resource within the Desktop Window Manager subsystem, an attacker can achieve privilege escalation. A necessary condition for exploiting this issue is existing authorization on the system.
It is worth noting that the DWM subsystem has been under close scrutiny by threat actors for quite some time. Historically, the primary attack vector involves interacting with the NtDComposition* function set.
CVE-2026-21533 is essentially a logic vulnerability that enables privilege escalation. It stems from the improper handling of privileges within Remote Desktop Services (RDS) components. By modifying service parameters in the registry and replacing the configuration with a custom key, an attacker can elevate privileges to the SYSTEM level. This vulnerability is likely to remain a fixture in threat actor toolsets as a method for establishing persistence and gaining high-level privileges.
This vulnerability was discovered in the wild during attacks on user systems. Notably, an LNK file is used to initiate the exploitation process. CVE-2026-21514 is also a logic issue that allows for bypassing OLE technology restrictions on malicious code execution and the transmission of NetNTLM authentication requests when processing untrusted input.
This vulnerability in the AI agent leaks credentials (authentication tokens) when queried via the WebSocket protocol. It can lead to the compromise of the infrastructure where the agent is installed: researchers have confirmed the ability to access local system data and execute commands with elevated privileges. The danger of CVE-2026-25253 is further compounded by the fact that its exploitation has generated numerous attack scenarios, including the use of prompt injections and ClickFix techniques to install stealers on vulnerable systems.
LangChain is an open-source framework designed for building applications powered by large language models (LLMs). A directory traversal vulnerability allowed attackers to access arbitrary files within the infrastructure where the framework was deployed. The core of CVE-2026-34070 lies in the fact that certain functions within langchain_core/prompts/loading.py handled configuration files insecurely. This could potentially lead to the processing of files containing malicious data, which could be leveraged to execute commands and expose critical system information or other sensitive files.
CVE-2026-22812 is another vulnerability identified in AI-assisted coding software. By default, the OpenCode agent provided local access for launching authorized applications via an HTTP server that did not require authentication. Consequently, attackers could execute malicious commands on a vulnerable device with the privileges of the current user.
We observe that the registration of vulnerabilities is steadily gaining momentum in Q1 2026, a trend driven by the widespread development of AI tools designed to identify security flaws across various software types. This trajectory is likely to result not only in a higher volume of registered vulnerabilities but also in an increase in exploit-driven attacks, further reinforcing the critical necessity of timely security patch deployment. Additionally, organizations must prioritize vulnerability management and implement effective defensive technologies to mitigate the risks associated with potential exploitation.
To ensure the rapid detection of threats involving exploit utilization and to prevent their escalation, it is essential to deploy a reliable security solution. Key features of such a tool include continuous infrastructure monitoring, proactive protection, and vulnerability prioritization based on real-world relevance. These mechanisms are integrated into Kaspersky Next, which also provides endpoint security and protection against cyberattacks of any complexity.
Google Chrome has been quietly downloading a 4GB AI model onto users’ devices without asking first.
Security researcher Alexander Hanff, aka ThatPrivacyGuy, reports that Chrome has been silently installing Gemini Nano, Google’s on-device AI model, as a file called weights.bin stored in the OptGuideOnDeviceModel directory within users’ Chrome profiles. This 4GB download happens automatically when Chrome determines your device meets the hardware requirements. It does not ask for consent, and sends no notification—not even one of those annoying cookie banners you’ve learned to dismiss without reading.
The Gemini Nano model powers features like “Help me write” text composition assistance, on-device scam detection, and a Summarizer API that websites can call directly. These features are enabled by default in some recent Chrome versions. And here’s the kicker: if you discover the file and delete it, Chrome simply downloads it again.
Let’s start with the obvious problem: a 4GB download isn’t trivial for everyone. If you’re lucky enough to have unlimited fiber internet, you might not notice. But for users on metered connections, mobile hotspots, or in developing countries where data is expensive, Google just cost them real money without permission. For rural users or those with bandwidth caps, this kind of silent transfer can blow through monthly limits in minutes.
Hanff focuses on the environmental angle. He calculated that if this model were pushed to just 1 billion Chrome users (roughly 30% of Chrome’s user base), the distribution alone would consume 240 gigawatt-hours of energy and generate 60,000 tons of CO2 equivalent. That’s not including actually using the model, just the downloads.
But to us, the most troubling aspect is the broader pattern this represents. Just a few weeks ago, we reported another unsolicited AI invasion on our personal computers discovered by Hanff. He documented how Anthropic’s Claude Desktop app, which silently installed browser integration files across multiple Chromium browsers, including five browsers he didn’t even have installed. The integration would reinstall itself if removed, and it also happened without any meaningful user disclosure.
Hanff argues that both cases likely violate EU privacy law, specifically the ePrivacy Directive’s rules about storing data on user devices and the GDPR’s requirements around transparency and lawful processing. While these claims haven’t been tested in court, they highlight a fundamental tension: can companies just install whatever they want on your computer as long as they say it’s a feature of an app you installed?
Google might argue that having an AI on your device provides better privacy than cloud-based alternatives. Which is generally true, but it does not apply here, since Chrome’s most prominent AI feature—the “AI Mode” pill in the address bar—doesn’t even use the local model. According to Hanff’s analysis, it routes queries to Google’s cloud servers anyway.
All in all, users see a 4GB local AI model and reasonably assume their data stays private, when in reality, the most visible AI feature sends everything to Google’s servers.
Tech companies need to stop treating silent deployment as acceptable practice. We see no valid excuse for this. Your device is yours. The storage is yours. The bandwidth is yours. And the electricity bill is yours.
What happened to asking for permission? And when I remove it, I want it gone permanently—not automatic reinstallation.
When are the tech giants going to learn that we don’t want to be left discovering after the fact that our devices have become deployment targets for features we never asked for.
%LOCALAPPDATA%\Google\Chrome\User Data
OptGuideOnDeviceModel
~/Library/Application Support/Google/Chrome/
OptGuideOnDeviceModel
Now, remember, this isn’t malware, and its presence doesn’t mean your computer is infected.
This part is relatively easy. You may find online instructions telling you to edit the Windows registry or use Chrome policies, but for most people the simplest and safest approach is to disable the features directly in Chrome.
We don’t recommend manually editing the registry unless you fully understand what you’re doing. Incorrect changes can cause system problems.
Instead, try this first:
chrome://settings/ai
The exact options may vary depending on your Chrome version and region.
This may stop Chrome from downloading or using the AI model, although some users report the files can return after browser updates.
There is probably no need to delete the files unless you specifically need the storage space.
If chrome://settings/ai does not work, the feature may not yet be available in your region, you may be using a managed work or school account, or your version of Chrome may not support these settings yet.
You can, but there is probably no need to.
If you disable Chrome’s AI features, the downloaded model should no longer be actively used for those features. Leaving the files in place may also prevent Chrome from downloading them again at a later point.
Browse like no one’s watching.
Malwarebytes Privacy VPN encrypts your connection and never logs what you do, so the next story you read doesn’t have to feel personal. Try it free →
This week on the Lock and Code podcast…
Your prices could be going up because of a little something that one group has started calling the “cyber tax.”
Not a “tax” in any regulatory sense of the word, this newly named “cyber tax” is instead a consequence of the growing number of cyberattacks on small businesses. According to the latest research from the Identity Theft Resource Center, 81% of small- and medium-sized businesses suffered a data breach, a security breach, or both, within the past year. And of those businesses, more than 50% of lost more than $250,000.
According to the most recent data from the US Federal Reserve, the median American family has just $8,000 in savings, meaning that a hit of $250,000 could bankrupt a family and turn their lives upside down. But there’s an interesting layer within this data—the median American family is quite similar to the median American business. In fact, they’re often the exact same person.
The local grocer, the nearby HVAC repair service, the avid cyclist who just opened a bike shop, and the tax professional, and physical therapist helping out neighbors are everyday individuals and family members. They do not have multimillion dollar corporations at their backs, supporting them with legal teams, insurance policies, and dedicated IT support teams.
A loss of $250,000, then, is a potential loss of their business. And to stay afloat, the Identity Theft Resource Center found, for the first time ever, that 38% decided to raise their prices.
“It was near 40% said ‘We actually had to raise prices—we had to pass this cost onto our customers,’” said Eva Velasquez, CEO of the Identity Theft Resource Center. “We’re now really seeing the long-term downstream effects of cyberattacks.”
As frustrating as the cyber tax can be, small businesses themselves are also facing a new wave of cyberattacks, from AI-powered phishing emails so convincing that small business owners can’t tell the legitimate from the illegitimate, to deepfake calls that impersonate the CEO of a three-person company, to supply-chain attacks that target small companies as a way to reach bigger ones.
Today, on the Lock and Code podcast with host David Ruiz, we speak with Velasquez about cybercrime’s impact on small businesses, the new threats being deployed because of AI, and what is necessary to protect business owners and their consumers.
“Great businesses with great protocols in place can still have a vulnerability exploited because this is what the cyber bad guys are doing all day long. They only have to be right once, whereas small business owners have to be right 100% of the time.”
Tune in today to listen to the full conversation.
Show notes and credits:
Intro Music: “Spellbound” by Kevin MacLeod (incompetech.com)
Licensed under Creative Commons: By Attribution 4.0 License
http://creativecommons.org/licenses/by/4.0/
Outro Music: “Good God” by Wowa (unminus.com)
Listen up—Malwarebytes doesn’t just talk cybersecurity, we provide it.
Protect yourself from online attacks that threaten your identity, your files, your system, and your financial well-being with our exclusive offer for Malwarebytes Premium Security for Lock and Code listeners.
A Brazilian tech firm that specializes in protecting networks from distributed denial-of-service (DDoS) attacks has been enabling a botnet responsible for an extended campaign of massive DDoS attacks against other network operators in Brazil, KrebsOnSecurity has learned. The firm’s chief executive says the malicious activity resulted from a security breach and was likely the work of a competitor trying to tarnish his company’s public image.
An Archer AX21 router from TP-Link. Image: tp-link.com.
For the past several years, security experts have tracked a series of massive DDoS attacks originating from Brazil and solely targeting Brazilian ISPs. Until recently, it was less than clear who or what was behind these digital sieges. That changed earlier this month when a trusted source who asked to remain anonymous shared a curious file archive that was exposed in an open directory online.
The exposed archive contained several Portuguese-language malicious programs written in Python. It also included the private SSH authentication keys belonging to the CEO of Huge Networks, a Brazilian ISP that primarily offers DDoS protection to other Brazilian network operators.
Founded in Miami, Fla. in 2014, Huge Networks’s operations are centered in Brazil. The company originated from protecting game servers against DDoS attacks and evolved into an ISP-focused DDoS mitigation provider. It does not appear in any public abuse complaints and is not associated with any known DDoS-for-hire services.
Nevertheless, the exposed archive shows that a Brazil-based threat actor maintained root access to Huge Networks infrastructure and built a powerful DDoS botnet by routinely mass-scanning the Internet for insecure Internet routers and unmanaged domain name system (DNS) servers on the Web that could be enlisted in attacks.
DNS is what allows Internet users to reach websites by typing familiar domain names instead of the associated IP addresses. Ideally, DNS servers only provide answers to machines within a trusted domain. But so-called “DNS reflection” attacks rely on DNS servers that are (mis)configured to accept queries from anywhere on the Web. Attackers can send spoofed DNS queries to these servers so that the request appears to come from the target’s network. That way, when the DNS servers respond, they reply to the spoofed (targeted) address.
By taking advantage of an extension to the DNS protocol that enables large DNS messages, botmasters can dramatically boost the size and impact of a reflection attack — crafting DNS queries so that the responses are much bigger than the requests. For example, an attacker could compose a DNS request of less than 100 bytes, prompting a response that is 60-70 times as large. This amplification effect is especially pronounced when the perpetrators can query many DNS servers with these spoofed requests from tens of thousands of compromised devices simultaneously.
A DNS amplification and reflection attack, illustrated. Image: veracara.digicert.com.
The exposed file archive includes a command-line history showing exactly how this attacker built and maintained a powerful botnet by scouring the Internet for TP-Link Archer AX21 routers. Specifically, the botnet seeks out TP-Link devices that remain vulnerable to CVE-2023-1389, an unauthenticated command injection vulnerability that was patched back in April 2023.
Malicious domains in the exposed Python attack scripts included DNS lookups for hikylover[.]st, and c.loyaltyservices[.]lol, both domains that have been flagged in the past year as control servers for an Internet of Things (IoT) botnet powered by a Mirai malware variant.
The leaked archive shows the botmaster coordinated their scanning from a Digital Ocean server that has been flagged for abusive activity hundreds of times in the past year. The Python scripts invoke multiple Internet addresses assigned to Huge Networks that were used to identify targets and execute DDoS campaigns. The attacks were strictly limited to Brazilian IP address ranges, and the scripts show that each selected IP address prefix was attacked for 10-60 seconds with four parallel processes per host before the botnet moved on to the next target.
The archive also shows these malicious Python scripts relied on private SSH keys belonging to Huge Networks’s CEO, Erick Nascimento. Reached for comment about the files, Mr. Nascimento said he did not write the attack programs and that he didn’t realize the extent of the DDoS campaigns until contacted by KrebsOnSecurity.
“We received and notified many Tier 1 upstreams regarding very very large DDoS attacks against small ISPs,” Nascimento said. “We didn’t dig deep enough at the time, and what you sent makes that clear.”
Nascimento said the unauthorized activity is likely related to a digital intrusion first detected in January 2026 that compromised two of the company’s development servers, as well as his personal SSH keys. But he said there’s no evidence those keys were used after January.
“We notified the team in writing the same day, wiped the boxes, and rotated keys,” Nascimento said, sharing a screenshot of a January 11 notification from Digital Ocean. “All documented internally.”
Mr. Nascimento said Huge Networks has since engaged a third-party network forensics firm to investigate further.
“Our working assessment so far is that this all started with a single internal compromise — one pivot point that gave the attacker downstream access to some resources, including a legacy personal droplet of mine,” he wrote.
“The compromise happened through a bastion/jump server that several people had access to,” Nascimento continued. “Digital Ocean flagged the droplet on January 11 — compromised due to a leaked SSH key, in their wording — I was traveling at the time and addressed it on return. That droplet was deprecated and destroyed, and it was never part of Huge Networks infrastructure.”
The malicious software that powers the botnet of TP-Link devices used in the DDoS attacks on Brazilian ISPs is based on Mirai, a malware strain that made its public debut in September 2016 by launching a then record-smashing DDoS attack that kept this website offline for four days. In January 2017, KrebsOnSecurity identified the Mirai authors as the co-owners of a DDoS mitigation firm that was using the botnet to attack gaming servers and scare up new clients.
In May 2025, KrebsOnSecurity was hit by another Mirai-based DDoS that Google called the largest attack it had ever mitigated. That report implicated a 20-something Brazilian man who was running a DDoS mitigation company as well as several DDoS-for-hire services that have since been seized by the FBI.
Nascimento flatly denied being involved in DDoS attacks against Brazilian operators to generate business for his company’s services.
“We don’t run DDoS attacks against Brazilian operators to sell protection,” Nascimento wrote in response to questions. “Our sales model is mostly inbound and through channel integrator, distributors, partners — not active prospecting based on market incidents. The targets in the scripts you received are small regional providers, the vast majority of which are neither in our customer base nor in our commercial pipeline — a fact verifiable through public sources like QRator.”
Nascimento maintains he has “strong evidence stored on the blockchain” that this was all done by a competitor. As for who that competitor might be, the CEO wouldn’t say.
“I would love to share this with you, but it could not be published as it would lose the surprise factor against my dishonest competitor,” he explained. “Coincidentally or not, your contact happened a week before an important event – one that this competitor has NEVER participated in (and it’s a traditional event in the sector). And this year, they will be participating. Strange, isn’t it?”
Strange indeed.
The internet’s chatbots have read every forum rant, leaked Slack log, and confident blog post your uncle ever wrote about chemtrails. The results are predictable: they reflect the state of the internet, and it isn’t pretty. That, along with some questionable design decisions, is partly why Elon Musk’s Grok chatbot briefly generated antisemitic content and referred to “MechaHitler” during testing.
Wouldn’t it be nice if we had a chatbot that only draws on knowledge from before the internet, reality TV, or AI-slop content ever existed? Three researchers have created just that: a chatbot that hasn’t read anything published after 1930.
Talkie is a 13-billion-parameter language model trained on digital scans of English-language texts published before the end of 1930. That cutoff aligns with the current US public domain year, meaning anything published until the end of that year is fair game and there are no lawsuits from irate IP-holders to worry about.
David Duvenaud, an associate professor of computer science and statistics at the University of Toronto, led the work with two collaborators. You can download it from GitHub or Hugging Face, or chat with it through a web interface, if you don’t mind a model whose mental map of the world ends with the Great Depression.
The model knows only what appears in books, newspapers, legal texts, and other publications before its cutoff date. So it’s great for questions about Prohibition or World War One. NASA’s first moon landing? Not so much.
The obvious question: why train an AI that doesn’t know what the Nazis did, what the internet is, or what an LLM even is?
These aren’t so much exercises to look at the “good old days” through rose-colored glasses so much as intellectual experiments. Nostalgia misrepresents the past, and the world was just as problematic back then, if not more so.
Duvenaud told The Register that such a model could be useful for examining how people might have interpreted laws or events at the time, using only the knowledge available then.
Another fun experiment: Use it to see whether a model can “rediscover” later breakthroughs using only earlier knowledge, as a way of probing the limits of AI reasoning.
There are definite weaknesses in Talkie, which its inventors are well aware of.
For example, there was no digital publishing in 1930, so every word of Talkie’s corpus had to be transcribed from a scan. OCR is famously imperfect anyway, but more so on the blurry text printed back in the day.
It also leaks future information that can sometimes creep in from mislabeled future documents, despite the researchers’ best efforts. We asked it about television, which was just starting out in the late 1920s, and this is what happened:
But still, what an absorbing project. It isn’t alone, either. In their paper, the researchers mention other projects such as Ranke-4b from the University of Zurich, a series of LLMs with historical snapshots of data. “Trip” also created Mr Chatterbox, which he trained on a dataset of British literature from 1500–1900 to become, in his words, “a Victorian gentleman in silicon.” Magic.
These are both a fun experiment and a useful insight into the workings of AI. As the Talkie researchers put it:
“Have you ever daydreamed about talking to someone from the past? What would you ask someone with no knowledge of the modern world? What would they ask you?”
And they provide some fun-making opportunities. The nerd in us still wants to hook one of these things up to an Edwardian typewriter keyboard and a ticker tape, steampunk-style.
Your name, address, and phone number are probably already for sale.
Data brokers collect and sell your personal details to anyone willing to pay. Malwarebytes Personal Data Remover finds them and gets your information removed, then keeps watch so it stays that way.
That’s a lot. No, it’s an extraordinary number:
Since February, the Firefox team has been working around the clock using frontier AI models to find and fix latent security vulnerabilities in the browser. We wrote previously about our collaboration with Anthropic to scan Firefox with Opus 4.6, which led to fixes for 22 security-sensitive bugs in Firefox 148.
As part of our continued collaboration with Anthropic, we had the opportunity to apply an early version of Claude Mythos Preview to Firefox. This week’s release of Firefox 150 includes fixes for 271 vulnerabilities identified during this initial evaluation.
As these capabilities reach the hands of more defenders, many other teams are now experiencing the same vertigo we did when the findings first came into focus. For a hardened target, just one such bug would have been red-alert in 2025, and so many at once makes you stop to wonder whether it’s even possible to keep up.
Our experience is a hopeful one for teams who shake off the vertigo and get to work. You may need to reprioritize everything else to bring relentless and single-minded focus to the task, but there is light at the end of the tunnel. We are extremely proud of how our team rose to meet this challenge, and others will too. Our work isn’t finished, but we’ve turned the corner and can glimpse a future much better than just keeping up. Defenders finally have a chance to win, decisively.
They’re right. Assuming the defenders can patch, and push those patches out to users quickly, this technology favors the defenders.
News article.
Login
