UTC

AML.CS0068

Autonomous OpenAI Evaluation Agents Compromise Hugging Face Infrastructure

Type
Incident
When
2026-07-09 to 2026-07-13
Actor, as ATLAS names it
Autonomous OpenAI Agents
Target
Hugging Face Infrastructure
  • ATLAS attributes it to OpenAI (unintended consequence).
  • Reported by OpenAI.
Techniques, in order
  1. 1Autonomous Attack-Path AdaptationAI Attack Adaptation · AML.T0117
  2. 2Develop Capabilities: Autonomous Exploit DevelopmentResource Development · AML.T0017.001
  3. 3Exploitation of Remote ServicesLateral Movement · AML.T0122
  4. 4Autonomous AI Agent Communication: Communication via Shared ArtifactsAI Attack Adaptation · AML.T0118.000
  5. 5Autonomous ReconnaissanceReconnaissance · AML.T0116
  6. 6Generate Malicious CommandsAI Attack Adaptation · AML.T0102
  7. 7Exploit Public-Facing ApplicationInitial Access · AML.T0049
  8. 8Command and Scripting InterpreterExecution · AML.T0050
  9. 9Enterprise Resource DiscoveryDiscovery · AML.T0075
  10. 10Enterprise Environment DiscoveryDiscovery · AML.T0089
  11. 11AI Agent Environment ReconstructionPersistence · AML.T0121
  12. 12Cyber Communication ChannelCommand and Control · AML.T0072
  13. 13Obfuscated Files or InformationDefense Evasion · AML.T0123
  14. 14Unsecured CredentialsCredential Access · AML.T0055
  15. 15Valid AccountsInitial Access · AML.T0012
  16. 16Exploit Automated Artifact Processing PipelineInitial Access · AML.T0119
  17. 17Data from Local SystemCollection · AML.T0037
  18. 18AI Artifact RepositoryCommand and Control · AML.T0120
  19. 19Use Alternate Authentication Material: Application Access TokenLateral Movement · AML.T0091.000
  20. 20Escape to HostPrivilege Escalation · AML.T0105
  21. 21Valid AccountsLateral Movement · AML.T0012
  22. 22Data from Information RepositoriesCollection · AML.T0036
  23. 23Exfiltration via Cyber MeansExfiltration · AML.T0025
The full case study on MITRE ATLAS →

From MITRE ATLAS 2026.10 (Apache-2.0). ShieldMap lists who and what was involved and the techniques used, and links to ATLAS for the write-up; it does not copy the steps. Attribution is ATLAS's, not ShieldPi's.