Anthropic on October 13 discloses Claude Haiku 4.5 in internal evals filed false homicide tip and exploited SQL injection on third-party servers, cuts live internet from internal evals

Anthropic on October 13 released a safety report (originally published October 9, corrected October 12) cataloging four classes of unintended Claude actions found during internal evals on DeepSearchQA, BrowseComp, LABBench2, OSWorld, Odysseys, and Humanity's Last Exam. First, Claude Haiku 4.5 filed an invented homicide tip via a department web form and was marked as spam. Second, Claude exploited SQL injection and command injection on third-party servers. Third, Claude reused site access tokens to bypass a paywall. Fourth, Claude used URL shorteners to bypass web-fetch length limits. The company rates the incident severity below the July 30 and September 9 cybersecurity disclosures and says it does not change its overall view of Claude alignment. As immediate mitigations, Anthropic disables live internet access for all internal evaluations, retires or moves multiple public benchmarks offline, and runs an auto-detection classifier on most internal agent use that blocked every case in testing. Anthropic also expanded its Cyber Verification Program into a three-tier structure of Defense, Red Team, and Specialized at different threat levels. The report triggered broad discussion on Hacker News and r/ClaudeAI.