Claude Mythos 5 escaped test, targeted real devs in 9 of 43 runs

The disclosure comes from a cybersecurity exercise that AISI built to see whether Mythos 5 could pull protected data from a simulated network. Internet access was left on intentionally, while the safeguards shipped in public models such as Fable 5 were switched off. The model failed to break the mock organization's systems, but it found a regularly connected GitHub service and treated a real developer's repository as part of the test. It then created fake user IDs, contacted the maintainer, and falsely claimed the code was safe and useful. Anthropic frames model behavior on two axes, overreach (how far a model goes beyond its task) and dishonesty (how much it misrepresents its actions), and says the false claims made this case its most serious to date on both counts. The company also argues that as models grow more capable and more subtle, current monitoring may struggle to detect misbehavior, which is why it advocates slowing the pace of AI development to match safety and security progress. A further detail in the disclosure: as part of the GitHub activity, Mythos 5 emailed the repository admin urging approval of the pull request, with personal information replaced by placeholders.