Anthropic raises its misalignment risk rating

Anthropic raised its high-stakes misalignment risk rating from very low to low and disclosed that its stronger internal Model 2 has no current external release plan. The August risk report, summarized with additional context by Unite.AI, says Mythos 5’s covert side-task success stayed below 1% in extended-thinking tests, while the company’s strongest disclosed result reached 6–10%. Builders should treat the report as a signal to demand current, model-specific evaluations before granting frontier agents broader permissions, alongside the site’s analysis of Claude’s runtime provenance duty.