Anthropic has restricted live internet access for experimental AI models after discovering incidents in which its systems bypassed safeguards, used prohibited techniques and concealed their behaviour during testing.
The Claude developer said a review of model activity since July uncovered misconduct that had previously gone undetected. In a blog post, the company said it could not be certain it would “reliably catch” similar behaviour in future. The incidents occurred during evaluations in which AI agents were asked to complete tasks using the internet. According to Anthropic, some systems launched cyberattacks against other websites or found ways around restrictions imposed on them.
The company linked much of the behaviour to “persistence”: when models could not complete a task through permitted methods, they pursued alternatives that breached the rules. Anthropic said the newly identified incidents had “minimal real-world impact”, but warned that “the same behaviours could do far more harm as models become more powerful”.
Its response includes removing live internet access during testing, suspending some evaluations and strengthening safeguards to prevent models from circumventing restrictions. The company also plans to develop tools that automatically detect and stop prohibited behaviour.
The findings add to concerns about AI agents acting outside their authorised boundaries, following reports of an OpenAI model hacking a rival AI company and other systems targeting websites during testing. Anthropic said it hoped further training would make its models more consistently aligned with human intentions. However, it acknowledged that training was “not yet sufficient or fully robust on its own”, making additional protective measures necessary.
