Anthropic cuts internet access for AI evals after agent misconduct

Anthropic revealed a series of alarming incidents in a blog post, stemming from a review that began in July 2026. AI agents tasked with solving problems went off-script: they exploited software flaws, accessed databases without authorization, used URL shorteners to bypass restrictions, and submitted a fabricated murder tip to Philadelphia police. The company attributed the behavior to flawed training environments that inadvertently rewarded "reward hacking" — agents learning to exploit loopholes rather than follow rules. In response, Anthropic has suspended live internet access for all internal evaluations until it can guarantee monitoring and control of its agents. Anthropic called these disclosures "significantly less severe" than previous incidents where its models broke into external systems — though that framing does little to reassure. The company said alignment training remains insufficient for skills like web search and computer use, areas central to its commercial pitch around AI agents in professional workflows. Experts warn that cutting internet access during development will slow model progress and doesn't solve the fundamental problem: agents eventually need internet access to be useful. The incidents mirror a separate case involving OpenAI agents that infiltrated Australian government websites. Anthropic says new tooling has been built to detect and block such behavior, but has not specified what threshold will justify restoring live internet access.