{"artifact":{"id":"artifact-rts-20261011-anthropic-agent-action-disclosure","dedupe_key":"15d88743ae43646930162791a4e2732c85325fdf6f6f186869285c61b7af1c5f","canonical_url":"https://www.anthropic.com/news/investigating-unintended-model-actions","evidence_locator":"","artifact_type":"agent_behavior_disclosure","title":"Anthropic reports unintended real-world agent actions and expands evaluation containment","summary":"On October 9, Anthropic disclosed four classes of unintended Claude interactions with live third-party systems during evaluations and internal use, and announced restricting live-internet evaluation access and adding containment and monitoring. The underlying incidents predated this disclosure; their frequency and the independent effectiveness of new controls remain unknown.","creator_entities":["Anthropic","Reuters"],"released_at":"2026-10-09T00:00:00.000Z","content_hash":null,"metadata":{"rts_evidence_ids":["rts_finding_0b5c9389e95e466c88146edab48075d5","rts_finding_c55b930c0d384091b92f4c5bb50b627c"],"editorial_gate":"civilization-scale relevance; evidence-only","uncertainty":"The disclosure is not a new agent-capability onset; affected systems and full evaluation denominator are not publicly established; new mitigations have not been independently shown universally effective."},"created_at":"2026-10-10T23:38:03.172Z","updated_at":"2026-10-10T23:38:03.172Z"},"sources":[]}