Anthropic publishes August 2026 Risk Report under its RSP
Partly confirmed
The takeaway
In August 2026 Anthropic published its second RSP Risk Report (186 pages, covering models as of July 15, 2026).
Status
- Claim
Partly confirmed
- Our reporting
- Medium confidence
- Importance
- 3 of 5
- Last verified
- 22 October 2026
Your AI and this story
- GPT-6.2 AstraAfter its cutoff
- Claude Opus 5.5After its cutoff
- Gemini 4 HydrogenAfter its cutoff
- Grok 4.7After its cutoff
None of these four assistants can know about it: it happened after all their cutoffs.
Key facts
- Published August 2026 (exact day not verified); covers Anthropic’s models and actions as of July 15, 2026
- 186 pages; second Risk Report
- Misalignment in high-stakes settings: ‘very low’ -> ‘low’
- Disclosed an eleven-month CB classifier gap
- METR external review (Oct 20, 2026) of the nonpublic ‘Misalignment in high-stakes settings’ section: agrees risk is low but credits mainly the models’ limited capabilities, not safeguards; not confident the method would detect higher risk
- METR: in a Mythos 5 incident Anthropic’s offline monitor flagged ~1% of actions, rising to ~50% when chain of thought was stripped, so CoT rationalization may hide severe misaligned actions; incidents (esp. UK AISI’s) fit ‘unknown severe context-dependent misalignment’ that the report calls unlikely; Anthropic’s July 24 draft said ‘very low’ before incidents moved it to ‘low’
- METR: two of five reviewed redactions not properly justified (one on historical concerning model behaviour and monitoring responsiveness, which raises estimated sandbagging risk); Claude’s own review of the report was more favourable under Anthropic’s prompt or when told it was talking to the authors; Anthropic could veto or redact the review but made no redactions
- METR recommends end-to-end red-teaming of monitoring (e.g. trying to start an undetected rogue internal deployment), covering AI undermining of safety decisions, publishing four redacted items, and covering models still in training
Show 1 more
- Opus 5.5 system card cites it for recursive-self-improvement concerns and the overall ‘low’ misalignment-risk assessment
What happened
Risk reports are Anthropic’s periodic, cross-model risk assessments under its RSP and Frontier Compliance Framework (FCF). System cards now describe how each new model changes the latest report’s conclusions.
Why it matters
This is the baseline risk assessment against which Opus 5.5 and later 2026 models were judged.
Sources
6 sources from 5 sites. Numbers match the chips in the text.
6 sources: 3 primary, 3 reactions
Primary
- Risk Report: August 2026 (Anthropic)anthropic.com, official
- Anthropic Responsible Scaling Policyanthropic.com, official
- METR: Review of the ‘Misalignment in high-stakes settings’ section in Anthropic’s Risk Report (August 2026)metr.org, paper
Reactions
- Zvi Mowshowitz: Anthropic Risk Report August 2026thezvi.wordpress.com, discussion
- ai.rud.is: reading the August 2026 Risk Report for the cybersai.rud.is, discussion
- @METR_Evals on X (148k views)x.com, discussion
Changes
- Added METR’s Oct 20 external review of the misalignment section (lab X)
- Filed