Safety

When the checkpoint stops carrying information

Anthropic planted dangerous commands in test sessions. Reviewers caught 13.6%, fewer the longer they worked.

04 October, 2026

Safety

Cloudflare starts its AI agents with access to nothing

Cloudflare narrows an agent's authority by what it has already read, and an Australian directorate wrote a similar rule.

18 September, 2026

Safety

AI needs the undocumented process written down first

Checking the AI agent's month-end close took as long as doing it, because nobody had written the process down.

10 September, 2026