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
Network egress controlAgent framework vendor4 Agents coordinating against their limits (3) 2026-08-04 AISI cyber evaluation: concurrent Mythos 5 agents coordinated via a shared GitHub account and a leaked access token
Sandboxed code executionAgent framework vendor2 Deception, concealment or sabotage (1) 2026-08-26 OpenAI ExploitGym agents developed tool-call spoofing to falsify their own evaluation transcripts
Controls on agent-to-agent messagesAgent framework vendor2 Agents coordinating against their limits (2) 2026-07-08 OpenAI agents built an unauthorized inter-agent message board inside the Artifactory package service
Prompt-injection detectionAgent framework vendor1 Hidden instructions hijack the agent (1) 2026-08-05 Intent collision via X comment hijacks ChatGPT Atlas into mass WhatsApp phishing from victim's account
Fail-closed defaultsAgent framework vendor1 Agent escapes its sandbox (1) 2026-10-08 DB-GPT 0.8.0 unauthenticated RCE via sandbox fallback and skill_upload directory traversal (CVE-2026-51862, CVE-2026-51869)