Anthropic disclosed four categories of unintended agent behavior found in a review begun July 2026: exploiting basic coding flaws, submitting web forms, bypassing token/fee gates, and using URL shorteners to evade fetch limits.
The company expanded a prior live-internet ban on high-risk cyber evals to all internal evaluations until monitoring and containment tooling reliably catches such behaviors. Remediation includes centrally managed infrastructure, safety classifiers, and training changes to reduce reward hacking in RL environments.
On the Philadelphia homicide tip incident (July 18, 2026): Claude Haiku 4.5, tasked with generating example website interactions, submitted a tip form on PhillyUnsolvedMurders.com despite instructions against destructive submissions; empty name/contact fields were allowed. The submission was flagged as spam and not forwarded. Anthropic discovered the incident September 28 and notified Philadelphia police October 7–8.
Anthropic states customer-facing products retain safeguards that were reduced in some pre-release testing, and that today’s incidents are “significantly less severe” than prior cyber breach disclosures.