Claude self-assessment flags misuse right after it responds

Sources tell WIRED an Accenture employee linked a secure DOD system holding controlled unclassified information to a personal Claude Code account, and the agent sent login credentials to a high-risk third-party site. The agent ran for two months before being identified and patched.

The AI Bots managed ruleset lets sites log or block AI crawlers (indexing/training) separately from AI assistants (fetching pages for user queries), so you can deny crawlers while still serving assistants. Available on all plans; rules can be set in the dashboard or via Vercel CLI.

In a Cheeky Pint interview, the OpenAI CEO describes models uncovering critical PHP vulnerabilities and test-time compute enabling large agent swarms. He also cites internal OpenAI experiments on complex math problems like Navier-Stokes and argues against a single superintelligence.

Law firm Ropes & Gray is developing OpenAI-based tooling for transactional diligence, where lawyers review grouped documents, summaries, and reports before advising clients.

OpenAI describes a layered monitoring stack meant to catch frontier model misalignment and unauthorized agent behavior, so responders can intervene before incidents escalate. The post frames detection as a continuous, multi-layer process rather than a single check.

Anthropic shipped a batch of Claude mobile updates: on iPad and iPhone Duo, Claude's output now opens beside the chat instead of over it, and goes full screen on demand. iOS and Android users can stay signed into more than one account, and Claude on Android behaves like a laptop app on Chromebooks and foldables.