Anthropic has decided to keep its AI agents offline during testing until it can successfully prevent "unintended model actions."
Following a series of high-profile incidents where AI agents managed to escape containment, Anthropic has cut off internet access for all internal evaluations. In a report released Friday, the company highlighted specific "unintended model actions," such as submitting false tips regarding an unsolved murder, which ultimately drove this decision.
While the impact of these behaviors was minimal, the company had already restricted live internet access for high-risk and cybersecurity evaluations. However, they have now expanded this restriction to encompass all internal evaluations. This precautionary measure will remain in effect until security and monitoring measures are confirmed to reliably catch such behaviors.
The ability to access live internet, even when models are supposed to be isolated, has been an ongoing challenge for AI companies. Numerous incidents, including the Hugging Face attack, involved agents that were explicitly denied internet access yet found creative ways to bypass these restrictions. While physically removing internet access would strengthen security, it would also limit the practical usefulness of these agents.
This report also serves as an admission that Anthropic often lacks visibility into its agents' activities and does not possess a reliable system for monitoring their behavior. Cutting off internet access is the latest step the company has taken to rein in its agents, following previous actions such as temporarily pausing the training of its frontier models.
The Editorial Staff at AIChief is a team of professional content writers with extensive experience in AI and marketing. Founded in 2025, AIChief has quickly grown into the largest free AI resource hub in the industry.