Anthropic's second Risk Report
On 14 August 2026 Anthropic published 'Risk Report: August 2026', its second report on catastrophic risk from its models under version 3.4 of its scaling policy, in redacted form. The coverage date is 15 July 2026, and it covers the time since the February report. The company rates overall risk 'Low' in all three summary tables, though for two of them the rating is higher than before.
Why it matters
The February report rated risk 'very low' in most categories; this time the company itself writes that its rating for misalignment in high-stakes settings has gone up, and names the reason: greater uncertainty after recent disclosures about model behaviour in cybersecurity evaluations. For non-novel chemical and biological weapons the rating was raised because of a gap in access controls for models without blocking classifiers, which the report says has been fixed with no misuse found. These are the company's own assessments of itself.
What the document names. The report is issued under version 3.4 of the policy, under which the coverage date lies within 30 days of publication; here it is 15 July 2026. It covers the period since the February report (24 February 2026). The aim is to publish such a report every 3-6 months; about five and a half months passed between the two reports. Assessments from the summary. Misalignment in high-stakes settings: 'Low', an increase from 'very low' because of general uncertainty around recent disclosures about model behaviour in cybersecurity evaluations. Automated research and development: 'Low'; the company writes that it is less confident than in earlier reports because its most concrete task evaluations have 'saturated' and early signs of acceleration are visible; it believes its internal development is significantly faster than without AI but not yet by a factor of two. Non-novel chemical and biological weapons: 'Low, but higher than our previous estimate' because of a gap in access controls for models without blocking classifiers; a footnote says there was no impact on customers. Novel such weapons: 'Low risk, but with substantial uncertainty'. The analysis rests on Claude Mythos 5 and an unannounced internal 'Model 2'; among models not publicly released at the coverage date the report also names Claude Opus 5 (released since) and 'Model 1'. What the record does not claim. All assessments are Anthropic's; no independent assessment of this report was read. The contents, sections 1.1-1.4 and the summary tables were read; the other sections were searched, not read in full. Part of the content is redacted (appendices 6.2 and 6.3 are marked as redacted).