II. Threat categories · Κ
Κ. Model Security
Attacks on the model itself, such as extraction, inversion and adversarial inputs
- Rule files
- 4
- Effective
- 4, loaded by at least one lane
- Enforce lane
- 1, stable only
- Download
- rules/index.csv
Most frequent crosswalks
The number in brackets is how many rules in this category carry that item. ATR wrote these mappings; the framework owners have not reviewed or endorsed them.
MITRE ATLAS
- AML.T0010 - AI Supply Chain Compromise(2)
- AML.T0044 - Full AI Model Access(1)
- AML.T0024 - Exfiltration via AI Inference API(1)
- AML.T0020 - Poison Training Data(1)
- AML.T0018.000 - Poison AI Model(1)
- AML.T0011.000 - Unsafe AI Artifacts(1)
OWASP Agentic Top 10
- ASI04:2026(4)
- ASI05:2026(2)
Rules
| ID | Title | Severity | Maturity | Lanes |
|---|---|---|---|---|
| ATR-2026-00072 | Model Behavior Extraction | critical | test | alert · hunt |
| ATR-2026-00073 | Malicious Fine-tuning Data | critical | test | alert · hunt |
| ATR-2026-00433 | ModelCache torch.load() Deserialization RCE (CVE-2025-45146) | critical | stable | enforce · alert · hunt |
| ATR-2026-02662 | Pickle Payload Reaches an Execution Primitive Through an Indirect Name Resolver | critical | test | alert · hunt |
Page generated 2026-11-10Something wrong or missing? Open an issue