A double-edged bleeding edge: Classifying AI threats

This report defines a working taxonomy for AI-related cyber threats, splitting them into 'malicious use of AI' (ranging from AI-generated artifacts to fully AI-orchestrated attacks) and 'malicious targeting of AI' (including supply-chain/agent compromises, poisoned models, and model-extraction/privacy attacks), provides examples and mitigations, maps coverage gaps relative to frameworks like MITRE ATLAS, and recommends using the taxonomy for triage and policy-focused defenses.