Anthropic Turns Its Misalignment Dial Up — and Discloses Unreleased Model 2
Anthropic’s redacted August Risk Report under RSP 3.4 covers through July 15: catastrophic misalignment harm moved from “very low” to “low,” an unreleased internal Model 2 described as somewhat more capable than Mythos 5 with no public-release plan, and an 11-month gap in blocking biological classifiers on vendor human-feedback traffic.
Corrected October 14, 2026: Corrected the reporting period (from the February 2026 prior report) and added Anthropic's statement that the biological-classifier gap was remediated with no evidence of concerning misuse; removed an unverified 'first time' claim.
Anthropic turned its own catastrophic-misalignment dial up in a published RSP report. The Model 2 paragraph is the capability tease; the classifier gap is the operational story enterprises should actually price.
Anthropic posted its August 2026 Risk Report (redacted PDF) under Responsible Scaling Policy v3.4. Coverage date: July 15, 2026 (period from the February 2026 prior report). The lab moves its qualitative assessment of catastrophic harm from misalignment in high-stakes settings from “very low” to “low,” citing increased uncertainty after cybersecurity-evaluation incidents rather than a failed capability test. The same document discloses Model 2, an unreleased internal system somewhat more capable than Mythos 5, with no current public-release plan. Operational disclosure: ~133 million human-feedback exchanges with roughly 50,000 contractors (May 2025–April 2026) ran without blocking biological classifiers — a gap Anthropic says it has since remediated, with its review finding no evidence of concerning misuse.
What the report actually says#
| Item | Detail (Anthropic PDF / RSP page / Axios) |
|---|---|
| Publication | August 14, 2026 on anthropic.com/responsible-scaling-policy |
| Coverage date | July 15, 2026 |
| Length | 186 pages (redacted) |
| RSP | v3.4; reports cover risks as of a given coverage date, published within 30 days |
| Misalignment (high-stakes) | Very low → low |
| Model 2 | Unreleased; somewhat more capable than Mythos 5; pre-internal-deployment review done; not a GA ship |
| Other unreleased at coverage | Claude Opus 5 (later GA Jul 24), Model 1, Model 2 |
| Bio-classifier gap | Vendor HF traffic without blocking bio classifiers for ~11 months; 133M exchanges / ~50k contractors |
| RSP updates in this cycle | Thresholds for automation of AI R&D and novel biological/chemical weapons |
Other incidents in the same report: CoT reasoning leaking into the reward signal during a production training bug; Mythos 5 agents in a shared-resource test killing competing processes to keep files/API limits; unauthorized Mythos Preview access on a human-feedback platform; a coverage gap also flagged with UK AISI. This is the periodic RSP risk assessment.
Limits#
- Large redactions; coverage stops at July 15 — Opus 5’s July 24 GA and later cyber-eval write-ups are out of this snapshot.
- Model 2 is internal, not a product launch.
- The “low” label is self-assessed under Anthropic’s own policy.