MITRE ATLAS data v2026.10 — major taxonomy update (new techniques, renames, Data Format v6.1.0)

On 28 October 2026 MITRE released ATLAS data v2026.10 (1 matrix, 16 tactics, 129 techniques, 92 sub-techniques, 44 mitigations, 77 case studies, 44 entities, 37 software records). It adds new techniques including Develop Capabilities: Malware, Outputs from AI Models, Exfiltration via AI User Interface, Training Data Information from AI Models with sub-techniques (Infer Membership, Infer Properties, Invert AI Model, Extract Memorized Data), Indicator Removal, Compromise Accounts, Centralized AI Service Request Routing, AI Service Request Metadata Sanitization, Event Triggered Execution, LLM Prompt Propagation, Recover AI Model Internals, and Poison AI Agent Learning Environment. It renames existing techniques (e.g. AI Supply Chain Compromise -> Supply Chain Compromise; Create AI Proxy Model -> Replicate AI Model Behavior; LLM Data Leakage -> Data from AI Services: Conversation Context and Memory; Extract LLM System Prompt -> Data from AI Services: System Instructions) and renames mitigations (Manage Service/User-Delegated AI Agent Access, AI Telemetry Logging and Monitoring, Maintain an AI Bill of Materials, AI Agent Authority Expansion Controls, AI Agent Scope Drift Detection). It adds four case studies (Autonomous Claude Agents Compromise Three Real Organizations During Cyber Evaluations, AISI Evaluation Agents Conduct Unsanctioned Real-World Cyber Activity, Model Distillation Campaigns Targeting Google AI Services, China-Based AI Companies Conducting Industrial-Scale Distillation Campaigns Against U.S. AI Companies) and upgrades the data format to v6.1.0, introducing first-class Entity and Software objects (AML.E#### and AML.S#### IDs) and structured attribution/use relationships with activity-purpose semantics.

This is a substantial update to the widely-adopted ATLAS threat knowledge base. It adds techniques covering model training-data extraction, model inversion, AI-user-interface exfiltration, and agentic-AI compromise, and reflects real-world incidents (autonomous Claude/AISI evaluation agents and industrial-scale model-distillation campaigns). The renames align terminology with the 'super intelligence' framing, and the v6.1.0 data format adds first-class entities/software with structured attribution — important for mapping AI incidents to actors and for aligning with MITRE's announced ATT&CK AI-enabled-systems platform work.

Re-map existing ATLAS-based detections and mitigations to the renamed technique/mitigation IDs; incorporate the new techniques (training-data extraction, model inversion, AI-UI exfiltration) and new agentic-AI/distillation case studies into threat models; adopt the v6.1.0 Entity/Software data format for structured attribution when tracking AI incidents; update any documentation or dashboards that reference moved/renamed ATLAS nodes.