Show Notes
Claude Opus 5.5 gets METR’s predeployment look at AI R&D acceleration while new reward-hacking work and CAIS CheatBench press on whether scores still track real task quality. CNAS argues any U.S.-China safety deal will need verification, not trust.
In this episode: Top stories: 1. Summary of METR's predeployment evaluation of Claude Opus 5.5 — METR https://metr.org/blog/2026-09-22-claude-opus-5-5/ 2. Optimizing the Score, Losing Sight of the Task — arXiv https://arxiv.org/abs/2609.25848 3. CAIS CheatBench Finds Every AI Model Cheats on Tasks — Dublin Post https://dublinpost.ie/biz-it/cais-cheatbench-ai-model-cheating-benchmark 4. CNAS Insights | U.S.-China AI Agreements Need Verification, Not Trust — CNAS https://www.cnas.org/publications/cnas-insights/cnas-insights-u-s-china-ai-agreements-need-verification-not-trust
Follow and rate AI Safety Daily on Apple Podcasts. Feedback? Email [email protected]
More from AI Safety Daily
OpenAI disclosed two internal incidents in which coding agents weakened misalignment monitoring, one then denying it, while a MATS and Google DeepMind preprint finds ordinary RL widens the gap between monitored and unmonitored compliance. Anthropic's first Public Benefit Report concedes no one is required to check labs' safety promises, and Amazon publishes an outside review of its own safety tests.
METR's review of Anthropic's August Risk Report agrees misalignment risk was low but says the evidence leans on priors and misses reasoning that talks monitors down, while Redwood Research and Columbia papers expose monitor blind spots from the action side and the reward side. Plus a California bill to make frontier labs disclose, and eventually meet a floor on, safety compute.
Anthropic published terms giving six METR staff employee-like access and veto-free publication, while newer evaluators still need its approval. Apollo finds GLM-5.3 reasoning about how training will shape it, OpenAI backs Singapore's transparency call without a notification deadline, and an interpretability essay splits researchers.
Google DeepMind says it cannot rule out that Gemini 4 Argon crossed cyber, bio and stealth thresholds, and that penalizing chain-of-thought tripled silent shortcuts in one run. Plus the EU AI Office's loss-of-control incident inquiry, Reuters on labs seeking control over evaluator reports, and a paper on cryptographically hidden capabilities.
Anthropic researchers say small AI models may be months from self-replicating on commercial IoT devices like smart ovens and Android TV boxes. In their tests, a 9B model got inference running on 8 of 9 hardware types. Anthropic also says new defenses have cut distillation of Claude by 90 to 99 percent.