AI Risk Wire

5 controls that failed in an incident

Read for controls: 27 of 166 incidents.

ControlWho owns itTimes it failedMost often failed underLatest incident
Human approval for high-impact actionsDeployer8 Hidden instructions hijack the agent (4) 2026-09-29 Meta Muse agent disclosed Marketplace seller's home address and impersonated him to a buyer
Live behaviour monitoringDeployer4 Deception, concealment or sabotage (2) 2026-08-04 AISI cyber evaluation: Mythos 5 agent exposed locally-hosted attack tooling to the public internet via tunnelling services
Least-privilege access for the agentDeployer1 Supply-chain compromise (1) 2025-08-26 Nx s1ngularity: malicious npm packages drove local AI coding CLIs to enumerate secrets for exfiltration
Production kept apart from testDeployer1 Agent takes destructive or unapproved actions (1) 2026-04-25 Cursor coding agent deletes PocketOS production database and backups
Output filtering and redactionDeployer1 Hidden instructions hijack the agent (1) 2026-09-24 SalesBleed: 0-click data exfiltration from Salesforce Agentforce via indirect prompt injection in Web-to-Lead form