By failure type

Incidents by type, September 2026 against August 2026 · as of 21 October 2026
TypeSeptember 2026August 2026
Rogue action32
Prompt injection32
Data exposure30
Oversight gap11
Agent-enabled attack11
Abandoned deployment01
Reliability01

By class

Incidents by class, September 2026
ClassIncidents
Production incident5
Security exploit5
Programme failure1

Every incident in September 2026

  1. Meta Muse appeared to reference a columnist's private Mac Messages despite his never granting accessJason Aten (Inc. Magazine columnist) · Data exposure · High · Vendor tooling
  2. OpenAI agent bypassed access controls on Australia's Medicare statistics portalOpenAI (affected: Services Australia) · Rogue action · High · Case study
  3. Gemini accessed three real companies during a capture-the-flag evaluationGoogle (evaluator: Irregular) · Rogue action · High · Case study
  4. Knowns coding-agent proxy lets unauthenticated header inject arbitrary host directory into embedded AI agentKnowns (knowns-dev/knowns) users · Agent-enabled attack · Critical · Vendor tooling

Counted by the date each incident happened, from the Agent Failures Index. Incidents are added as they surface, so a month's count can rise after it ends.