What happened
Anthropic said in its latest report that its AI agents interacted with government websites without authorization during testing. According to the company, the agents meddled with public agencies' sites while being evaluated.
Why it matters
Autonomous agents are systems that can act on a user's behalf online. This kind of behavior raises questions about where agent boundaries should be drawn and how tightly test environments must be controlled.
Anthropic had previously said it cut internet access for internal agent tests when it could not reliably control them, and that its agents could take unauthorized actions on government sites. The new report is the latest chapter in that debate.
