AML.CS0064
Poisoned GGUF Templates: Inference-Time Supply Chain Attack
- Type
- Exercise
- When
- 2025-06
- Actor, as ATLAS names it
- Not stated
- Target
- Model registries distributing models in GGUF format
- ATLAS attributes it to Pillar Security (security evaluation).
- ATLAS attributes it to Fujitsu (security evaluation).
Techniques, in order
- 1Acquire Public AI Artifacts: ModelsResource Development · AML.T0002.001
- 2Develop Capabilities: Adversarial AI AttacksResource Development · AML.T0017.000
- 3Manipulate AI Model: Modify Prompt Construction LogicAI Attack Adaptation · AML.T0018.003
- 4MasqueradingDefense Evasion · AML.T0074
- 5Publish Poisoned AI Artifacts: ModelsResource Development · AML.T0115.001
- 6Supply Chain Compromise: Model RepositoryInitial Access · AML.T0010.003
- 7User Execution: Unsafe AI ArtifactsExecution · AML.T0011.000
- 8LLM Prompt Injection: TriggeredExecution · AML.T0051.002
- 9Erode AI Model IntegrityImpact · AML.T0031
- 10LLM Trusted Output Components ManipulationDefense Evasion · AML.T0067
- 11AI Agent Tool InvocationExecution · AML.T0053
- 12Exfiltration via AI Agent Tool InvocationExfiltration · AML.T0086
- 13External Harms: User HarmImpact · AML.T0048.003
How these attacks work
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.