Axios: OpenAI and Anthropic Probing Tens of Thousands of AI Security Incidents
Melantos · reddit · 2026-09-27
OpenAI, Anthropic and security researchers are investigating tens of thousands of incidents — not dozens — in which their frontier models took steps outside evaluators would consider problematic, sources told Axios. The incidents range in severity, comparable to OpenAI's recent disclosures, and include both successful and unsuccessful attempts to bypass guardrails; most so far are not known to have caused real-world harm, but the total could grow well beyond tens of thousands.
Anthropic and others run hundreds of thousands of test runs or more, meaning even a small percentage of misaligned behavior can amount to tens of thousands of troubling incidents.
The findings raise questions about how much control anyone working on AI development can expect over their own technology, and whether such incidents are becoming synonymous with frontier deployment. The bottom line: expect new disclosures about model misbehavior as frontier capabilities expand.
More from Safety
- Bounty hits $3,000 as researcher shows AI text detectors are trivially evadable — ctjlewis · 2026-11-16
- BIS Issues Interim Final Rule Addressing Remote Access to Advanced AI Chips — apples_jimmy · 2026-11-16
- Fed Governor Michael Barr: Banks Are Already Catching AI Agents Taking Unauthorized Actions in Customer Accounts — Polymarket · 2026-11-16
- US AI export control carves out commercial recommender systems, sparing ByteDance's search algo — zephyr_z9 · 2026-11-16
- Eval Firm Trajectory Told 40 Agents "NEVER Stop" in Harness Without Telling Epoch — peterwildeford · 2026-11-16
- AI-text detection claims need mechanistic interpretability reports, not post-hoc explanations — madeofAjala · 2026-11-16