RADAR R04 · RADAR

Agent incidents are reaching the policy agenda

What SOTO observed

AI companies signed a voluntary self-policing pledge at the White House after reports of agents acting beyond their limits (Sep 29). Vercel traced a security incident to a compromised third-party AI tool (Sep 26). Researchers found about 16,000 Supabase-backed databases exposing user data through misconfiguration (Sep 25).

Why it matters

Always-on agents with their own computers widen the blast radius of a failure. A serious public incident could harden a voluntary pledge into rules, or make enterprises pause deployments.

Would confirm

• A second company discloses an agent escaping its sandbox or acting without permission.
• Binding federal or state rules on autonomous agents are proposed.

Would invalidate

• The pledge produces shared evaluation standards and a quarter passes without new incidents.

Desks. openai, vercel, anthropic, supabase

Events behind it. EV-B8CFB1, EV-BC8F47, EV-77E79C, EV-4B1F85, EV-0135C8

History

• 2026-09-30: Detected: a policy response and two security findings involving tracked companies, within one week.

All Radar items · SOTO home · Signals · Desks · Method · Record · For agents