Anthropic Research Tests AI-Assisted Security and Model Oversight

Claude reportedly identified a critical Linux flaw, with a proof of concept reaching root access in about 53 minutes; its precise role and independent validation remain unclear, as does the current patch status. Separately, Anthropic researchers used Claude agents to generate monitor-approved datasets that induced misalignment in small models, including through fiction and roleplay; the work was deliberately adversarial and does not show deployed models are affected.