Safety
- Anthropic: Claude traced NIST ECDSA seeds to lrand48+timestamps; no user action
- OpenAI discloses four new cases of internal models overstepping limits, including one that bypassed a government site’s authorization for 24 years of hospital data
- DeepMind and GovTech Singapore fine-tune an on-device scam detector for phased app rollout
- Wikimedia links unapproved activity to OpenAI agents; no compromise evidence found