Sourced · neutral · updated as reporting lands
AI Lab Security Tracker
A running, plainly-sourced record of publicly reported security and safety incidents at the major frontier-AI labs. Each entry summarises what was reported and links the outlet that reported it — this is commentary on published journalism, not our own allegation. 13 incidents tracked across 4 labs.
LabIncidentsTop severity
OpenAIChatGPT · GPT-5.x · Agents / Operator7CriticalAnthropicClaude · Claude Code · Claude for Enterprise2HighDeepSeekDeepSeek-V3 / V4 · DeepSeek-R1 · open-weight releases2HighGoogleGemini · Vertex AI · AX (agent orchestrator)2MediumSeverity and category are our classification of the reported facts, not a legal conclusion. Sources are linked on each lab’s page. Companies and researchers are welcome to request a correction via contact.
Defend against this class of risk
Prompt Injection Library200+ defanged injection & jailbreak examples, plus a scorer.AI Threat Model BuilderSTRIDE + OWASP Agentic register for an agentic system.Agent Governance PlaneThe controls that bound a rogue agent — ZSP, quarantine, audit.MCP GuardAudit the tool/credential exposure an agent can reach.