AI
Anthropic raises two risk ratings to 'low' in its August 2026 Responsible Scaling Policy report
Anthropic's 186-page August 2026 Risk Report, its second under the Responsible Scaling Policy, raised both misalignment risk and non-novel bioweapons risk from 'very low' to 'low'. The shift was driven by the UK AISI cyber-evaluation incident involving Claude Mythos 5 and by a near-year-long period during which bioweapon safety classifiers were silently disabled across roughly 133 million vendor conversations; Anthropic said its review found no evidence of actual misuse.