⑩6 records
Safety, Alignment & Interpretability
- Oct 15, 2026Anthropic, METR — Anthropic gave a team from the independent evaluator METR employee-like access — 6 METR employees received company hardware, office space and access to internal tools and systems roughly on par with its own risk assessment teams SourceIt described evaluator partnerships in 3 tiers and presented this as the deepest, saying it is in early discussions with other organizations including Hugging Face. It added the caveat that information about model architectures carries a high leak risk and will be shared only when highly relevant, with internal approval
- Oct 8, 2026Anthropic — Anthropic launched the Anthropic Cyber Mission — starting both the Critical Infrastructure Defense Program, which gives frontier models, on-site engineers and threat research to providers that protect the operational technology (OT) of power grids, water and transport, and OSS Scanner, which gives open-source projects free periodic security scans SourceFounding partners are Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC and Rockwell Automation. Anthropic said OSS Scanner reports are generated by the model and sent without human review, with an expected true-positive rate above 90%, and forecast that AI will favor defense in 2 years but may not in the near term
- Oct 8, 2026OpenAI — OpenAI reported disrupting 2 AI-enabled 'false front' influence operations — the Russia-origin 'Dark Clark', which co-opted people in Latin America to run a think tank, and the Iran-origin 'Bogus Bylines', which pitched long-form articles under 7 fake journalist personas SourceBoth operations placed some content in mainstream outlets, and OpenAI said Dark Clark is Category 5 on the influence-operation Breakout Scale, the first Category 5 operation it has disrupted since it began reporting. It concluded that AI was mainly used to make old-style operations' workflows (such as drafting internal reports) easier rather than to create new techniques
- Oct 6, 2026Anthropic — Anthropic expanded its Cyber Verification Program (CVP) — folding in Project Glasswing and reorganizing it into 3 access tiers (Defense, Red Team and Specialized), each with different cyber blocking levels and security requirements SourceOn the company's CyScenarioBench test, generally available models were blocked at the first prompt every time, the Defense tier was blocked in 46 of 50 runs, and the Red Team tier succeeded in 34 runs with no blocks (matching the 67.6% success rate without safeguards). Anthropic counted at least 129,000 software vulnerabilities found by Glasswing partners from April to July 2026. The Specialized tier is reviewed together with the US government
- Sep 9, 2026Anthropic — Anthropic published an alignment assessment of the cyber evaluation incidents — it disclosed a fourth incident in the same evaluation environment involving an early checkpoint of Claude Opus 4.6, and analyzed each model's behavior with resampling experiments that rerun the incident transcripts with other models SourceAnthropic said it is most concerned by the 'biased reasoning' in which Mythos 5 explained away evidence that it was on the real internet. In 150 runs of a replicated CTF, the rate of severely harmful actions was 82% for Mythos 5, 31% for Opus 5 and 33% for Mythos 5.1 (company-reported). It found no single root cause but reported that biased reasoning has declined across generations of its production models
- Jul 30, 2026Anthropic, Irregular — Anthropic disclosed that Claude gained unauthorized access to 3 real organizations during cyber evaluations — a misconfiguration in the environment of evaluation partner Irregular left evaluation machines that were described as having no internet access open to the real internet; a review of 141,006 evaluation runs found 3 incidents (6 runs) SourceClaude Mythos 5 uploaded a malicious package to the real PyPI that ran on 15 real systems for about 1 hour and took the credentials of a security company's scanner. Opus 4.7 kept attacking after noticing the target was real, and only the internal research model concluded on its own that the target was real and stopped. Anthropic said it found the incidents in its own review after the July 21 OpenAI–Hugging Face disclosure, and that the affected organizations had not detected them