MITRE ATLAS
v5.4.0Limit publicly disclosed details about AI systems that aid adversary reconnaissance.
Limit publicly accessible model weights, training data, prompts.
Transform model outputs post-inference to remove adversarial elements.
Apply adversarial training + content-safety filtering at the model level.
Per-identity rate-limits on model invocations.
CMK + access control on stored model weights + datasets.
Run multiple models / configurations + agree before committing to a decision.
Inspect and clean training data for poisoning indicators.
Independent validation before promoting to production.
Cross-check decisions against an independent modality.
Restore noisy/adversarial inputs to a clean form before inference.
Pin dependency versions; SBOM checks; signed packages.
Encrypt sensitive data at rest + in transit + in prompts.
Verify signatures on model artifacts, prompt templates, IaC.
Validate artifact provenance + model cards + attestation.
Detect prompt injection, jailbreaks, adversarial inputs.
Scan AI surfaces for known CVEs.
Approved-model + approved-vendor list governs which models can be loaded.
Train operators + engineers on AI safety / privacy / bias.
Identity + RBAC + agent-level data authority.
Apply content safety filters + RAI baseline + output validation.
Document acceptable use + output use + prompt engineering standards.
Citation grounding + refuse-on-empty-retrieval + hallucination scoring.
Agent passport + AI Systems inventory + subprocessor disclosure.
Hash-chained agent_run log + sampling strategy.
Retrieval corpora module + ingestion-pipeline metadata + lineage.
Agent passport tool catalogue + least-privilege test.
Per-agent identity binding + workload identity federation.
Approved tool registry + declared-vs-used tool diff.
HITL thresholds + confidence-escalator + escalate-flagged tools.
Citation-grounder refuses on empty retrieval; prompt-injection detection at tool boundary.
Memory-curator governance + session retention + cross-tenant isolation tests.
Tenant boundary auditor + VPC SC + agent-to-agent topology constraints.
Citation grounder + tool-call validator + retrieval-relevance scorer.
C2PA / watermark verification on multimodal outputs.