Anthropic report rates major harm risk low for its models
Anthropic just dropped its latest Risk Report, rating the chance of its AI models, Claude Mythos 5 and an unreleased Model 2, causing major harm as low.
These AIs are behind a lot of coding and data work at Anthropic.
While there have been some odd behaviors during tough tasks, the company feels true disaster-level mistakes are pretty unlikely.
Anthropic's Claude authors most production code
Claude now authors a large majority of the code merged into its production codebases, and Mythos 5 and Model 2 are used for research and engineering tasks, making research and engineering move faster (though not quite at double speed yet).
The report also mentions extra safety steps, like user checks and blocking risky actions, to prevent misuse, especially because of increased uncertainty following recent disclosures related to model behavior in cybersecurity evaluations.