Anthropic August 2026 Risk Report: Risk Still Low

Anthropic’s August 2026 Risk Report — its second public risk report — assesses the risk of catastrophic harm from its models as “low but not negligible,” and it contains a disclosure most labs would bury: biological classifiers were not running on human-feedback vendor traffic for roughly a year. The report was published August 14 under version 3.4 of Anthropic’s Responsible Scaling Policy, with a coverage date of July 15, 2026.

The report covers the period from the February 24, 2026 report through July 15, and it is the first under the RSP’s updated rules, which now require risk reports to cover risks “as of a given coverage date within 30 days of publication.”

Key facts

• Catastrophic misalignment risk: low but not negligible — Anthropic also revised its February 2026 assessment from “very low” to “low” in light of new information.
• Biological classifier gap disclosed: from May 2025 (first deployment of CB safeguards) until April 2026, all traffic collecting human-feedback data from contractor-evaluators ran without biological classifiers; fully remediated in April 2026, no evidence of concerning CB misuse, and no customer data or model weights exposed.
• RSP v3.4 threshold updates — updated thresholds for automation of AI R&D and for development of novel biological and chemical weapons, plus new coverage-date and external-review rules.
• Unreleased frontier models as of the coverage date: Claude Opus 5 (internally deployed, publicly released July 24) and “Model 1,” broadly similar in capability to the Mythos line, with limited internal deployment.
• Third-party red-teaming: a METR staff member spent three weeks red-teaming Anthropic’s automated offline monitoring pipeline in February–March 2026, finding several novel vulnerabilities, some since patched, none as severe as known ones.

What changed

The report’s structure is the change. RSP 3.4 formalizes what a risk report must cover — risks as of a coverage date, with external-review options for unredacted versions — and Anthropic’s own assessment moved misalignment risk up one notch, from “very low” to “low,” based on new information learned during the coverage period.

The bio-classifier disclosure is the most consequential finding for observers. The gap ran from May 2025 to April 2026 on vendor-evaluation traffic — a year in which Anthropic was shipping its biology safeguards and publishing system cards claiming coverage. Anthropic says the issue is fully remediated and that it now runs classifiers on the vast majority of vendor traffic, but the report itself concedes the discovery “leads us to believe that there is an increased likelihood of other, similar issues unknown to us.”

This is not the first time Anthropic’s safety apparatus has faced scrutiny this summer — the company investigated three real-world incidents in its cybersecurity evaluations in July, a story we covered when the findings landed.

Why it matters

• “Low” is a new data point for the whole industry. Anthropic is the only lab publishing periodic, RSP-graded risk reports. The deliberate one-notch escalation — and the explicit “not negligible” qualifier — gives regulators and enterprises a concrete reference level for frontier-model risk language.
• Safety claims now come with audit trails. The bio-classifier gap shows the difference between published safeguards and operational reality. Enterprise buyers evaluating Claude’s enterprise deployments get a public record of how a frontier lab found, disclosed, and remediated a coverage failure.
• The RSP itself is becoming a template. Between the updated thresholds, external review, and coverage-date rules, Anthropic’s framework increasingly resembles the kind of mandatory regime policymakers have proposed — and it is being voluntarily, publicly tested.

What to watch

• The weights window and cyber findings in follow-up model documentation, especially for any model flagged near the automation-of-AI-R&D threshold.
• External reviews. The report notes pilot reviews with METR and SecureBio; whether the LTBT exercises its new power to request formal external review will be a governance test.
• Independent verification of the bio-classifier fix, and whether the next report shows a clean coverage period.

Official source

• Anthropic: Redacted Risk Report — August 2026 (published August 14, 2026)
• Anthropic: Responsible Scaling Policy — current and prior versions

Updated August 16, 2026.