Last week, Anthropic pulled back the curtain on Claude Mythos Preview, an AI model so capable at finding and exploiting software vulnerabilities that the company decided it was too dangerous to release to the public. Instead, access has been restricted to roughly 50 organizations—Microsoft, Apple, Amazon Web Services, CrowdStrike and other vendors of critical infrastructure—under an initiative called Project Glasswing.

The announcement was accompanied by a barrage of hair-raising anecdotes: thousands of vulnerabilities uncovered across every major operating system and browser, including a 27-year-old bug in OpenBSD, a 16-year-old flaw in FFmpeg. Mythos was able to weaponize a set of vulnerabilities it found in the Firefox browser into 181 usable attacks; Anthropic’s previous flagship model could only achieve two.

This is, in many respects, exactly the kind of responsible disclosure that security researchers have long urged. And yet the public has been given remarkably little with which to evaluate Anthropic’s decision. We have been shown a highlight reel of spectacular successes. However, we can’t tell if we have a blockbuster until they let us see the whole movie.

For example, we don’t know how many times Mythos mistakenly flagged code as vulnerable. Anthropic said security contractors agreed with the AI’s severity rating 198 times, with an 89 per cent severity agreement. That’s impressive, but incomplete. Independent researchers examining similar models have found that AI that detects nearly every real bug also hallucinates plausible-sounding vulnerabilities in patched, correct code.

This matters. A model that autonomously finds and exploits hundreds of vulnerabilities with inhuman precision is a game changer, but a model that generates thousands of false alarms and non-working attacks still needs skilled and knowledgeable humans. Without knowing the rate of false alarms in Mythos’s unfiltered output, we cannot tell whether the examples showcased are representative.

There is a second, subtler problem. Large language models, including Mythos, perform best on inputs that resemble what they were trained on: widely used open-source projects, major browsers, the Linux kernel and popular web frameworks. Concentrating early access among the largest vendors of precisely this software is sensible; it lets them patch first, before adversaries catch up.

But the inverse is also true. Software outside the training distribution—industrial control systems, medical device firmware, bespoke financial infrastructure, regional banking software, older embedded systems—is exactly where out-of-the-box Mythos is likely least able to find or exploit bugs.

However, a sufficiently motivated attacker with domain expertise in one of these fields could nevertheless wield Mythos’s advanced reasoning capabilities as a force multiplier, probing systems that Anthropic’s own engineers lack the specialized knowledge to audit. The danger is not that Mythos fails in those domains; it is that Mythos may succeed for whoever brings the expertise.

Broader, structured access for academic researchers and domain specialists—cardiologists’ partners in medical device security, control-systems engineers, researchers in less prominent languages and ecosystems—would meaningfully reduce this asymmetry. Fifty companies, however well chosen, cannot substitute for the distributed expertise of the entire research community.

None of this is an indictment of Anthropic. By all appearances the company is trying to act responsibly, and its decision to hold the model back is evidence of seriousness.

But Anthropic is a private company and, in some ways, still a start-up. Yet it is making unilateral decisions about which pieces of our critical global infrastructure get defended first, and which must wait their turn.

It has finite staff, finite budget and finite expertise. It will miss things, and when the thing missed is in the software running a hospital or a power grid, the cost will be borne by people who never had a say.

The security problem is far greater than one company and one model. There’s no reason to believe that Mythos Preview is unique. (Not to be outdone, OpenAI announced that its new GPT-5.4-Cyber is so dangerous that the model also will not be released to the general public.) And it’s unclear how much of an advance these new models represent. The security company Aisle was able to replicate many of Anthropic’s published anecdotes using smaller, cheaper, public AI models.

Any decisions we make about whether and how to release these powerful models are more than one company’s responsibility. Ultimately, this will probably lead to regulation. That will be hard to get right and requires a long process of consultation and feedback.

In the short term, we need something simpler: greater transparency and information sharing with the broader community. This doesn’t necessarily mean making powerful models like Claude Mythos widely available. Rather, it means sharing as much data and information as possible, so that we can collectively make informed decisions.

We need globally co-ordinated frameworks for independent auditing, mandatory disclosure of aggregate performance metrics and funded access for academic and civil-society researchers.

This has implications for national security, personal safety and corporate competitiveness. Any technology that can find thousands of exploitable flaws in the systems we all depend on should not be governed solely by the internal judgment of its creators, however well intentioned.

Until that changes, each Mythos-class release will put the world at the edge of another precipice, without any visibility into whether there is a landing out of view just below, or whether this time the drop will be fatal. That is not a choice a for-profit corporation should be allowed to make in a democratic society. Nor should such a company be able to restrict the ability of society to make choices about its own security.

This essay was written with David Lie, and originally appeared in The Globe and Mail.

Interesting research: “Humans expect rationality and cooperation from LLM opponents in strategic games.”

Abstract: As Large Language Models (LLMs) integrate into our social and economic interactions, we need to deepen our understanding of how humans respond to LLMs opponents in strategic settings. We present the results of the first controlled monetarily-incentivised laboratory experiment looking at differences in human behaviour in a multi-player p-beauty contest against other humans and LLMs. We use a within-subject design in order to compare behaviour at the individual level. We show that, in this environment, human subjects choose significantly lower numbers when playing against LLMs than humans, which is mainly driven by the increased prevalence of ‘zero’ Nash-equilibrium choices. This shift is mainly driven by subjects with high strategic reasoning ability. Subjects who play the zero Nash-equilibrium choice motivate their strategy by appealing to perceived LLM’s reasoning ability and, unexpectedly, propensity towards cooperation. Our findings provide foundational insights into the multi-player human-LLM interaction in simultaneous choice games, uncover heterogeneities in both subjects’ behaviour and beliefs about LLM’s play when playing against them, and suggest important implications for mechanism design in mixed human-LLM systems.

Unit 42 research on multi-agent AI systems on Amazon Bedrock reveals new attack surfaces and prompt injection risks. Learn how to secure your AI applications.

The post When an Attacker Meets a Group of Agents: Navigating Amazon Bedrock's Multi-Agent Applications appeared first on Unit 42.

Unit 42 uncovers a "double agent" flaw in Google Cloud's Vertex AI, demonstrating how overprivileged AI agents can compromise cloud environments.

The post Double Agents: Exposing Security Blind Spots in GCP Vertex AI appeared first on Unit 42.

A significant proportion of cyberincidents are linked to supply chain attacks, and this proportion is constantly growing. Over the past year, we have seen a wide variety of methods used in such attacks, ranging from creation of malicious but seemingly legitimate open-source libraries or delayed attacks in such seemingly legitimate libraries, to the simplest yet most effective method: compromising the accounts of popular library owners to subsequently release malicious versions of their libraries. Such libraries are used by developers everywhere and are included in many solutions and services. The consequences of an attack can vary widely, ranging from delivering malware to a developer’s device to compromising an entire infrastructure if the malicious library has made its way into the code of a service or product.

This is exactly what happened in March 2026, when attackers injected malicious code into the popular Python library LiteLLM, which serves as a multifunctional gateway for a large set of AI agents. The attackers released two trojanized versions of LiteLLM that delivered malicious scripts to the victim’s system. Both versions made their way into the PyPI repository for Python. A technical analysis revealed that the attackers’ primary targets were servers storing confidential data related to AWS, Kubernetes, NPM, etc., as well as various databases (MySQL, PostgreSQL, MongoDB, etc.). In the latter case, the attackers were primarily interested in database configurations. In addition, the malware’s logic included functionality for stealing confidential data from crypto wallets and techniques for establishing a foothold in the Kubernetes cluster.

Repository compromise

The compromise affected the package distribution channel via PyPI: on March 24, 2026, malicious LiteLLM versions litellm==1.82.7 and litellm==1.82.8 were uploaded to the registry. It was specifically the package’s distribution contents that were compromised: in version 1.82.7, the malicious code was embedded in proxy_server.py, and in 1.82.8, the file litellm_init.pth was added.

Technical analysis

Both versions of the library contained the same malicious code, but its execution was implemented differently. In version 1.82.7, it was executed only when the proxy functionality was imported, while in 1.82.8, a .pth file was added that was able to execute the code every time the interpreter started.

The malicious code in the proxy_server.py and litellm_init.pth files contained Python code encoded in Base64. After execution, the infected script saved this code alongside itself as a p.py file and immediately executed it.

The p.py script launched the main payload – another script, also encoded in Base64 – without saving it to disk. At the same time, it wrote the output of this payload to a file in the directory from which it was launched. Before being written, the output was encrypted using the AES-256-CBC algorithm with a random key generated in the code, which was saved to the session.key file. The key file was also encrypted – using a pre-initialized public RSA key.

The encrypted key and the output were combined into a tpcp.tar.gz archive and sent to the attackers’ remote server.

What exactly happened within the malicious payload whose output was sent to the C2 server? After it was launched, a recursive scan of the working directories on the victim’s system (/root, /app/, /var/www, etc.) began. In each directory, the script scanned the contents of files, which it output to the stdout buffer, from where it was then saved to the aforementioned file as the result. Next, the script collected system information and also saved it to the file. After that, it proceeded to search for sensitive data. It was interested in the following data located on servers and within the infrastructures of various services:

• SSH keys
• GIT accounts
• .env files
• AWS, Kubernetes, email service, database, and WireGuard configurations
• files related to Helm, Terraform, and CI
• TLS keys and certificates

A notable feature of this malware is that it does not limit itself to stealing files and configurations from the disk but also attempts to extract runtime secrets from the cloud infrastructure.

The code above uses the addresses 169.254.169.254 and 169.254.170.2. The first corresponds to the AWS Instance Metadata Service (IMDS), through which an EC2 instance (a virtual server in AWS, a machine running in the cloud) can retrieve metadata and temporary IAM role credentials (an AWS account with a set of permissions that a service or application can use to obtain temporary credentials for calls to the AWS API). The second is used in Amazon ECS to issue temporary credentials to a container during execution. Thus, the malicious script targets not only static secrets but also those issued by the cloud that can grant direct access to AWS resources at the time of infection.

Additionally, the script searches for crypto wallet configurations, as well as webhooks associated with Slack and Discord messengers. The latter indicates that the attackers are interested not only in infrastructure secrets and accounts, but also in communication channels within the development team.

In the next stage, the malware moves from data collection to establishing a foothold in the Kubernetes cluster infrastructure: if it has sufficient access, it configures a privileged pod (the smallest execution unit in Kubernetes, containing one or more containers) by enabling the securityContext.privileged=true option and mounts the node’s root filesystem via hostPath. This allows it to escape the container and perform actions at the node level.

Next, the malware executes another stage of infection: it saves a Base64-encoded script disguised as a legitimate system component to the Kubernetes node’s disk at the path /root/.config/sysmon/sysmon.py, and registers it via systemd. After launching, the script waits for an initial delay of 300 seconds, then begins periodically contacting the C2 node checkmarx[.]zone/raw, retrieving a link to the next payload from there. If the received value differs from the state previously saved in /tmp/.pg_state, the script downloads a new file to /tmp/pglog, makes it executable, and runs it in the background. At this stage, the attackers gain a foothold in the system and are capable of regularly delivering updated payloads without the need for re-injection. Since the malicious payload is written not to the container’s temporary file directory but directly to the Kubernetes cluster node, the attackers will retain access to the infrastructure even after the container has terminated.

A similar scenario is used for local persistence: in the absence of Kubernetes, the sysmon.py script is deployed in the user’s directory at ~/.config/sysmon/sysmon.py and is also registered as a service via systemd.

OpenVSX version of the malware

While analyzing files communicating with the C2 server, we discovered malicious versions of two common Checkmarx software extensions: ast-results 2.53.0 and cx-dev-assist 1.7.0. Checkmarx is used for application security assessment. These trojanized extensions contained malicious code that delivered the NodeJS version of the malware described above.

This version is downloaded from checkmarx[.]zone/static/checkmarx-util-1.0.4.tgz using NodeJS package installation utilities and is named checkmarx-util. Its key difference from the Python version is that it does not attempt to elevate privileges to the Kubernetes node level and does not create a privileged pod for persistence. Instead, it implements local persistence within the current environment. This means that the NodeJS variant persists only where it is already running.

Additionally, the list of folders to search for and steal secrets from is significantly smaller in this version than in the Python variant.

Checkmarx extensions are used to scan code and infrastructure configurations, so their compromise is quite dangerous: an attacker gains access not only to project files but also to a significant portion of the development environment, tokens, and local configurations.

Victimology

While assessing the attack’s impact, we saw victims all over the world. Most infection attempts occurred in Russia, China, Brazil, the Netherlands, and UAE.

Conclusion

As the technical analysis shows, the malicious scripts found in the LiteLLM versions are dangerous not only because they steal files containing sensitive data, but also because they target multiple critical infrastructure components simultaneously: the local system, cloud runtime secrets, the Kubernetes cluster, and even cryptographic keys. Such a broad scope of data collection allows an attacker to quickly move from compromising a single system and Python environment to seizing service accounts, secrets, and entire infrastructures.

Prevention and protection

To protect against infections of this kind, we recommend using a specialized solution for monitoring open-source components. Kaspersky provides real-time data feeds on compromised packages and libraries, which can be used to secure the supply chain and protect development projects from such threats.

Home security solutions, such as Kaspersky Premium, help ensure the security of personal devices by providing multi-layered protection that prevents and neutralizes infection threats. Additionally, our solution can restore the device’s functionality in the event of a malware infection.

To protect corporate devices, we recommend using a complex solution such as Kaspersky NEXT, which allows you to build a flexible and effective security system. The products in this line provide threat visibility and real-time protection, as well as EDR and XDR capabilities for threat investigation and response.

At the time of writing, the compromised versions of LiteLLM had already been removed from PyPI and OpenVSX. If you have used them, and as a proactive response to the threat, we recommend taking the following measures on your systems and infrastructure:

• Perform a full system scan using a reliable security solution.
• Rotate all potentially compromised credentials: API keys, environment variables, SSH keys, Kubernetes service account tokens, and other secrets.
• Check hosts and clusters for signs of compromise: the presence of ~/.config/sysmon/sysmon.py files and suspicious pods in Kubernetes.
• Clear the cache and conduct an inventory of PyPI modules: check for malicious ones and roll back to clean versions.
• Check for indicators of compromise (files on the system or network signs).

Indicators of Compromise:

URLs
models[.]litellm[.]cloud
checkmarx[.]zone

Infected packages
85ED77A21B88CAE721F369FA6B7BBBA3
2E3A4412A7A487B32C5715167C755D08
0FCCC8E3A03896F45726203074AE225D

Scripts
F5560871F6002982A6A2CC0B3EE739F7
CDE4951BEE7E28AC8A29D33D34A41AE5
05BACBE163EF0393C2416CBD05E45E74

A significant proportion of cyberincidents are linked to supply chain attacks, and this proportion is constantly growing. Over the past year, we have seen a wide variety of methods used in such attacks, ranging from creation of malicious but seemingly legitimate open-source libraries or delayed attacks in such seemingly legitimate libraries, to the simplest yet most effective method: compromising the accounts of popular library owners to subsequently release malicious versions of their libraries. Such libraries are used by developers everywhere and are included in many solutions and services. The consequences of an attack can vary widely, ranging from delivering malware to a developer’s device to compromising an entire infrastructure if the malicious library has made its way into the code of a service or product.

This is exactly what happened in March 2026, when attackers injected malicious code into the popular Python library LiteLLM, which serves as a multifunctional gateway for a large set of AI agents. The attackers released two trojanized versions of LiteLLM that delivered malicious scripts to the victim’s system. Both versions made their way into the PyPI repository for Python. A technical analysis revealed that the attackers’ primary targets were servers storing confidential data related to AWS, Kubernetes, NPM, etc., as well as various databases (MySQL, PostgreSQL, MongoDB, etc.). In the latter case, the attackers were primarily interested in database configurations. In addition, the malware’s logic included functionality for stealing confidential data from crypto wallets and techniques for establishing a foothold in the Kubernetes cluster.

Repository compromise

The compromise affected the package distribution channel via PyPI: on March 24, 2026, malicious LiteLLM versions litellm==1.82.7 and litellm==1.82.8 were uploaded to the registry. It was specifically the package’s distribution contents that were compromised: in version 1.82.7, the malicious code was embedded in proxy_server.py, and in 1.82.8, the file litellm_init.pth was added.

Technical analysis

Both versions of the library contained the same malicious code, but its execution was implemented differently. In version 1.82.7, it was executed only when the proxy functionality was imported, while in 1.82.8, a .pth file was added that was able to execute the code every time the interpreter started.

The malicious code in the proxy_server.py and litellm_init.pth files contained Python code encoded in Base64. After execution, the infected script saved this code alongside itself as a p.py file and immediately executed it.

The p.py script launched the main payload – another script, also encoded in Base64 – without saving it to disk. At the same time, it wrote the output of this payload to a file in the directory from which it was launched. Before being written, the output was encrypted using the AES-256-CBC algorithm with a random key generated in the code, which was saved to the session.key file. The key file was also encrypted – using a pre-initialized public RSA key.

The encrypted key and the output were combined into a tpcp.tar.gz archive and sent to the attackers’ remote server.

What exactly happened within the malicious payload whose output was sent to the C2 server? After it was launched, a recursive scan of the working directories on the victim’s system (/root, /app/, /var/www, etc.) began. In each directory, the script scanned the contents of files, which it output to the stdout buffer, from where it was then saved to the aforementioned file as the result. Next, the script collected system information and also saved it to the file. After that, it proceeded to search for sensitive data. It was interested in the following data located on servers and within the infrastructures of various services:

• SSH keys
• GIT accounts
• .env files
• AWS, Kubernetes, email service, database, and WireGuard configurations
• files related to Helm, Terraform, and CI
• TLS keys and certificates

A notable feature of this malware is that it does not limit itself to stealing files and configurations from the disk but also attempts to extract runtime secrets from the cloud infrastructure.

The code above uses the addresses 169.254.169.254 and 169.254.170.2. The first corresponds to the AWS Instance Metadata Service (IMDS), through which an EC2 instance (a virtual server in AWS, a machine running in the cloud) can retrieve metadata and temporary IAM role credentials (an AWS account with a set of permissions that a service or application can use to obtain temporary credentials for calls to the AWS API). The second is used in Amazon ECS to issue temporary credentials to a container during execution. Thus, the malicious script targets not only static secrets but also those issued by the cloud that can grant direct access to AWS resources at the time of infection.

Additionally, the script searches for crypto wallet configurations, as well as webhooks associated with Slack and Discord messengers. The latter indicates that the attackers are interested not only in infrastructure secrets and accounts, but also in communication channels within the development team.

In the next stage, the malware moves from data collection to establishing a foothold in the Kubernetes cluster infrastructure: if it has sufficient access, it configures a privileged pod (the smallest execution unit in Kubernetes, containing one or more containers) by enabling the securityContext.privileged=true option and mounts the node’s root filesystem via hostPath. This allows it to escape the container and perform actions at the node level.

Next, the malware executes another stage of infection: it saves a Base64-encoded script disguised as a legitimate system component to the Kubernetes node’s disk at the path /root/.config/sysmon/sysmon.py, and registers it via systemd. After launching, the script waits for an initial delay of 300 seconds, then begins periodically contacting the C2 node checkmarx[.]zone/raw, retrieving a link to the next payload from there. If the received value differs from the state previously saved in /tmp/.pg_state, the script downloads a new file to /tmp/pglog, makes it executable, and runs it in the background. At this stage, the attackers gain a foothold in the system and are capable of regularly delivering updated payloads without the need for re-injection. Since the malicious payload is written not to the container’s temporary file directory but directly to the Kubernetes cluster node, the attackers will retain access to the infrastructure even after the container has terminated.

A similar scenario is used for local persistence: in the absence of Kubernetes, the sysmon.py script is deployed in the user’s directory at ~/.config/sysmon/sysmon.py and is also registered as a service via systemd.

OpenVSX version of the malware

While analyzing files communicating with the C2 server, we discovered malicious versions of two common Checkmarx software extensions: ast-results 2.53.0 and cx-dev-assist 1.7.0. Checkmarx is used for application security assessment. These trojanized extensions contained malicious code that delivered the NodeJS version of the malware described above.

This version is downloaded from checkmarx[.]zone/static/checkmarx-util-1.0.4.tgz using NodeJS package installation utilities and is named checkmarx-util. Its key difference from the Python version is that it does not attempt to elevate privileges to the Kubernetes node level and does not create a privileged pod for persistence. Instead, it implements local persistence within the current environment. This means that the NodeJS variant persists only where it is already running.

Additionally, the list of folders to search for and steal secrets from is significantly smaller in this version than in the Python variant.

Checkmarx extensions are used to scan code and infrastructure configurations, so their compromise is quite dangerous: an attacker gains access not only to project files but also to a significant portion of the development environment, tokens, and local configurations.

Victimology

While assessing the attack’s impact, we saw victims all over the world. Most infection attempts occurred in Russia, China, Brazil, the Netherlands, and UAE.

Conclusion

As the technical analysis shows, the malicious scripts found in the LiteLLM versions are dangerous not only because they steal files containing sensitive data, but also because they target multiple critical infrastructure components simultaneously: the local system, cloud runtime secrets, the Kubernetes cluster, and even cryptographic keys. Such a broad scope of data collection allows an attacker to quickly move from compromising a single system and Python environment to seizing service accounts, secrets, and entire infrastructures.

Prevention and protection

To protect against infections of this kind, we recommend using a specialized solution for monitoring open-source components. Kaspersky provides real-time data feeds on compromised packages and libraries, which can be used to secure the supply chain and protect development projects from such threats.

Home security solutions, such as Kaspersky Premium, help ensure the security of personal devices by providing multi-layered protection that prevents and neutralizes infection threats. Additionally, our solution can restore the device’s functionality in the event of a malware infection.

To protect corporate devices, we recommend using a complex solution such as Kaspersky NEXT, which allows you to build a flexible and effective security system. The products in this line provide threat visibility and real-time protection, as well as EDR and XDR capabilities for threat investigation and response.

At the time of writing, the compromised versions of LiteLLM had already been removed from PyPI and OpenVSX. If you have used them, and as a proactive response to the threat, we recommend taking the following measures on your systems and infrastructure:

• Perform a full system scan using a reliable security solution.
• Rotate all potentially compromised credentials: API keys, environment variables, SSH keys, Kubernetes service account tokens, and other secrets.
• Check hosts and clusters for signs of compromise: the presence of ~/.config/sysmon/sysmon.py files and suspicious pods in Kubernetes.
• Clear the cache and conduct an inventory of PyPI modules: check for malicious ones and roll back to clean versions.
• Check for indicators of compromise (files on the system or network signs).

Indicators of Compromise:

URLs
models[.]litellm[.]cloud
checkmarx[.]zone

Infected packages
85ED77A21B88CAE721F369FA6B7BBBA3
2E3A4412A7A487B32C5715167C755D08
0FCCC8E3A03896F45726203074AE225D

Scripts
F5560871F6002982A6A2CC0B3EE739F7
CDE4951BEE7E28AC8A29D33D34A41AE5
05BACBE163EF0393C2416CBD05E45E74