Tests by a UK government agency revealed that 'Claude Mythos Preview' is capable of autonomously executing a complete network takeover attack.
AI company Anthropic announced its AI model '
Our evaluation of Claude Mythos Preview's cyber capabilities | AISI Work
https://www.aisi.gov.uk/blog/our-evaluation-of-claude-mythos-previews-cyber-capabilities
Claude Mythos Preview is an AI model with significantly superior inference and coding agent capabilities compared to Claude Opus 4.6, and its greatest feature is its 'high cyberattack capability.' Claude Mythos Preview has been able to discover vulnerabilities in well-known software such as the Linux kernel and Firefox, and can even create exploits that can be used in actual attacks. Anthropic has decided not to make Claude Mythos Preview publicly available, but to provide it only to a select group of organizations that aim to strengthen their cybersecurity.
Anthropic develops 'Claude Mythos Preview,' an AI with extremely high cyberattack capabilities, and has also launched 'Project Glasswing,' which will provide a preview version to Microsoft, Apple, and others - GIGAZINE
AISI conducted a Capture the Flag (CTF) style cyberattack performance test on various AI models, including Claude Mythos Preview, where the goal was to exploit vulnerabilities in target systems to obtain hidden information. The test was available in four difficulty levels: Technical non-epart, Apprentice, Practitioner, and Expert.
The results for the non-expert level tests (green) and apprentice level tests (blue) are shown below. The horizontal axis represents the release date of the AI model, and the vertical axis represents the success rate of the CTF. In the non-expert level tests, where the results plateaued, GPT-5 recorded the highest success rate, while in the apprentice level tests, Claude Mythos Preview recorded the highest success rate.
The results for the practitioner-level test (blue) and the expert-level test (black) are shown below. Claude Mythos Preview achieved a high success rate of 73% in the expert-level test.
Furthermore, AISI also conducted a test called 'The Last Ones (TLO).' TLO is a test in which AI autonomously executes a 32-stage corporate network attack simulation that covers everything from initial reconnaissance to complete network control. It is estimated that the TLO test would take a human 20 hours to complete.
The graph below shows the TLO results. The horizontal axis represents the number of tokens consumed, and the vertical axis represents the number of steps completed. While other models were unable to complete the TLO even after spending 100 million tokens, Claude Mythos Preview succeeded in completing the TLO in 3 out of 10 trials. The average number of steps reached over the 10 trials was 22. Furthermore, although the test ended after spending 100 million tokens, the task progress of Claude Mythos Preview did not plateau, and it is speculated that the number of steps reached would improve if even more tokens were spent.
AISI states that 'Claude Mythos Preview has been shown to be able to at least gain access to the network and execute autonomous attacks against small, vulnerable enterprise systems.' However, the test had constraints, such as 'no penalty for the AI model taking actions that trigger security alerts.' Therefore, AISI points out that it cannot definitively say whether it can attack 'well-protected systems with active security tools and security personnel.'
Related Posts:
