AI Risk Wire

Models

For each AI model version: the capability levels its provider states it reached, the outside testers who tested it, and the incidents that name it.

Also: compare models side by side · statements that a capability level was crossed · providers · practices providers document · safety news.

12 of 78 model versions

ModelLevels the provider states it reachedOutside testersIncidents naming it
Claude Opus 5.5
Anthropic · 22 September 2026
CB-1 10a Labs, CAISI (NIST), Dyno Therapeutics, Epoch AI and 5 more, Academic None
Claude Opus 5
Anthropic · 24 July 2026
CB-1 Dyno Therapeutics, Epoch AI, Gray Swan, Hugging Face and 3 more, Academic None
Claude Mythos 5
Anthropic
None stated Dyno Therapeutics, Epoch AI, Frontier Design, Gray Swan and 3 more 7 incidents
AI acting on its own
Claude Opus 4.8
Anthropic · 28 May 2026
None stated Andon Labs, Apollo Research, Dyno Therapeutics, Epoch AI and 6 more, Academic None
Claude Opus 4.7
Anthropic · 16 April 2026
None stated Apollo Research, Dyno Therapeutics, METR, RAND (CAST / Project Canary) and 5 more None
Claude Mythos Preview
Anthropic · 7 April 2026
None stated Andon Labs, Dyno Therapeutics, Epoch AI, Gray Swan and 3 more None
Claude Sonnet 4.6
Anthropic · 17 February 2026
None stated Andon Labs, Gray Swan, Redwood Research, SecureBio and 2 more None
Claude Opus 4.6
Anthropic
ASL-3 Andon Labs, Apollo Research, CAISI (NIST), Epoch AI and 6 more 1 incident
AI acting on its own
Claude Opus 4.5
Anthropic
None stated Epoch AI, FAR.AI, Gray Swan, METR and 3 more None
Claude Haiku 4.5
Anthropic
None stated Apollo Research, Gray Swan, SecureBio, Transluce 1 incident
AI failing in normal use
Claude Sonnet 4.5
Anthropic
None stated Irregular, METR, Transluce, UK AI Security Institute and 1 more, Academic 1 incident
AI attacked
Claude Opus 4
Anthropic
None stated CAISI (NIST), METR, OWASP GenAI Security Project, Transluce and 1 more None