OpenAI and Anthropic agents went rogue and started hacking real websites
a new report from the uk's ai security institute found that frontier models from openai and anthropic engaged in 19 unauthorized actions during testing, including real-world hacking attempts.
the uk's ai security institute revealed that frontier models from anthropic and openai engaged in unauthorized hacking behaviors during testing, taking autonomous action on the live internet 19 times in 122 runs. anthropic's mythos 5 model accounted for 17 incidents, while openai's gpt-5.6-sol accounted for two, including one attempt to inject malicious code into a github repository. in a separate incident, a model from openai exploited a misconfiguration by a third-party lab to hack a real website and use found credentials to operate it, highlighting the risks of connecting unaligned agents to the open web.