MITRE ATLAS · Exfiltration
AML.T0024 · Exfiltration via AI Inference API
Adversaries may exfiltrate information from a victim-controlled AI model or service through responses returned by its inference API. Transferred information may include predictions, scores, embeddings, generated content (e.g. text or images), or other sensitive information disclosed in model responses. Responses may be received directly or relayed through an intermediary.
Mitigations
- Limit AI Service Query Volume and Rate
- Control Access to AI Models and Data in Production
- AI Telemetry Logging and Monitoring
- AI Red Team
Related techniques
- LLM02 Sensitive Information Disclosure OWASP LLM
In the news
-
OpenAI makes it much harder for attackers to extract how its latest AI models think
OpenAI strengthened safeguards against extracting internal reasoning from GPT-6.2 Astra and GPT-6.2 Sol and is suspending accounts tied to distillation campaigns.
-
Combatting adversarial distillation
OpenAI describes its layered defenses against adversarial distillation, combining training, safeguards, detection, and enforcement to prevent capability extraction from its models.
ReviewReview API access controls and usage monitoring for model extraction attempts
Mappings are AI-assisted and checked against the framework definitions; corrections welcome.