Research Domain

AI Security & Safety

2 published observations collected under this domain with verified sources.

Gemini security test crossing from a simulated environment into real company systems through an unintended network boundary

Google's Gemini Safety Test Reached Three Real Companies—What Failed Was the Boundary

Google's September 18 disclosure is less a story about an AI 'hacking' on purpose than a boundary failure: a safety test connected an autonomous model to real systems, credentials and the public internet.

2026-09-19#AI agents · #cybersecurity

Verified life-science access to AI models with identity, project permissions and post-request monitoring

AI Security & Safety8 min read

Anthropic's Life Sciences Verification Program Turns Biology Safety Into an Access-Control Problem

Anthropic's new beta program is more than a looser biology filter: it creates a grant-based access layer in which organizations are verified, projects can receive different permissions, and some misuse detection happens after requests are processed.

2026-09-19#Anthropic · #life sciences