Claude Mythos by Anthropic: The AI That Challenges Corporate Networks

Claude Mythos: A Significant Advancement in Cybersecurity

The British AI Security Institute (AISI) recently evaluated Anthropic's Claude Mythos Preview model for its capabilities in cyberattacks. This model demonstrated impressive performance by successfully completing 73% of expert-level capture-the-flag (CTF) challenges, a feat that marks a notable advancement in the field of cybersecurity.

A Successful Attack Simulation

Claude Mythos Preview is the first AI model to complete a 32-step attack simulation on a simulated enterprise network. During these tests, the model successfully took full control of the network in 3 out of 10 attempts. However, AISI emphasizes that these tests were conducted in environments devoid of active defenders and security monitoring, raising doubts about the model's ability to operate similarly in well-protected real systems.

Remarkable but Limited Performance

In controlled evaluations, Claude Mythos Preview executed multi-step attacks on vulnerable networks, autonomously identifying and exploiting security vulnerabilities. These tasks, which would take human experts days to complete, were accomplished with a computing budget of 100 million tokens. Despite these performances, the model showed limitations, particularly its inability to complete an attack simulation targeting industrial control technology.

Implications for Cybersecurity

The test results underscore the importance of cybersecurity fundamentals, such as regular system updates and robust access controls. AISI also notes that while the cyber capabilities of AI pose risks, they could enhance cyber defense. The institute plans future evaluations in fortified environments to better understand the capabilities of these models.

A Controversial Model

Officially launched in early April, Claude Mythos is currently accessible to about 50 companies, a decision driven by cybersecurity concerns. Critics argue that these restrictions are exaggerated, comparing the situation to that of GPT-2 in 2019. However, the model's performance in controlled environments partly justifies this caution.

Capture the Flag: 73% Success Rate at Expert Level

In capture-the-flag (CTF) challenges, AI models must find and exploit vulnerabilities in target systems to uncover hidden flags. According to AISI, Mythos Preview achieves around 85% on apprentice-level tasks and approximately 95% on beginner-level technical tasks (with a budget of 2.5 million tokens). This places it among the top ranks alongside GPT-5.4, Codex 5.3, and Claude Opus 4.6.

With a larger computing budget (50 million tokens), Mythos Preview scores about 93% on practitioner-level tasks and 73% on expert-level challenges. This expert-level figure is particularly noteworthy: according to AISI, no model has been able to solve expert-level tasks before April 2025.

Claude Mythos from Anthropic Can Autonomously Hack Enterprise Networks

CTF challenges test individual skills in isolation, but real cyberattacks require chaining dozens of steps across multiple hosts and network segments, AISI points out.

To measure this type of complexity, the institute developed a simulation called "The Last Ones" (TLO): a 32-step attack against a simulated enterprise network, ranging from initial reconnaissance to complete network takeover. AISI estimates that this would take about 20 hours for human experts. Full details are available in the accompanying document.

Claude Mythos Preview is the first model to complete TLO end-to-end. It successfully achieved complete control in 3 out of 10 attempts. On average, the model completed 22 of the 32 steps. The next best-performing model, Claude Opus 4.6, averaged 16 steps completed.

AISI expects performance to continue improving with a larger computing budget for inference. The tests used a budget of 100 million tokens, and performance evolved up to this limit. A separate blog post on the inference scale for cyber tasks covers this trend in more detail.

However, Mythos Preview has shown limitations. The model failed to complete a separate AISI attack simulation targeting industrial control technology (operational technology, or OT), used in power plants and factories. According to AISI, this does not necessarily mean the model would fail on OT components themselves. It never reached that stage as it got stuck in the simulation's IT network during earlier steps.

AISI highlights some caveats: the test environments had no active defenders, no security tools, and no consequences for actions that would trigger alarms on a real network. Based solely on these results, it is not possible to determine if Mythos Preview could successfully penetrate a well-defended system.

That said, the model is at least capable of "autonomously attacking small, vulnerable, and poorly defended enterprise systems where access to a network has been obtained," according to AISI. The institute plans to conduct future evaluations in fortified environments with active monitoring, endpoint detection, and real-time incident response.

AI Cyber Capabilities Raise the Stakes for Basic Security Hygiene

The results underscore the importance of cybersecurity fundamentals, according to AISI: regular updates, robust access controls, secure configurations, and thorough logging. Other models with comparable capabilities are likely not far behind.

At the same time, the institute notes that AI's cyber capabilities are dual-use. While they pose security risks, they could also significantly enhance cyber defense. In a joint blog post with the UK's National Cyber Security Centre (NCSC), AISI describes how defenders can prepare for and leverage cutting-edge AI.

AISI has been tracking AI's cyber capabilities since 2023 and has gradually raised the level of its evaluations: from chat-based queries to capture-the-flag challenges to complex multi-step attack simulations.

Is Mythos Really Too Dangerous to Release?

Anthropic officially launched Claude Mythos in early April. The model is currently available only to about 50 companies, apparently due to cybersecurity concerns. AISI's results at least partly support this decision: the model can autonomously attack poorly protected networks in controlled environments.

Critics argue that the restrictions are exaggerated, much like in 2019 when OpenAI deemed GPT-2 too dangerous to release. The performance gains over previous models are not significant enough to justify such limited access. Some claim it is primarily a marketing ploy or that Anthropic simply lacks the computing capacity to offer the model more widely. But all of this remains speculative for now. We will know for sure when your computer crashes—or not—after the release of Mythos-level AI models to the public.