currently_learning / ai-threat-frameworks
Two structured frameworks for AI threat and risk management.
MITRE ATLAS & NIST AI RMF
Structured frameworks for AI threat and risk management
These two frameworks answer different questions and pair well together. MITRE ATLAS catalogs how real attackers target AI systems — the specific tactics and techniques, the way ATT&CK does for traditional infrastructure. The NIST AI RMF answers the organizational question: how do you govern, measure, and manage AI risk across a system's lifecycle, regardless of the specific attack in play.
cat ./mitre-atlas.md
MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems) is a living knowledge base maintained by MITRE — the same organization behind ATT&CK — that documents real adversary tactics and techniques against AI and ML systems, built from actual incidents and red-team research rather than theory. Where ATT&CK covers general IT infrastructure, ATLAS is scoped specifically to the AI attack lifecycle, and its 2026 updates lean heavily into agentic AI: context and memory poisoning, agent configuration tampering, credential harvesting through tool calls, and exfiltration via agent-to-agent messaging.
Gathering intelligence on a target AI system, its model, data, and deployment.
Acquiring infrastructure, datasets, or tooling to stage an attack against the model.
Gaining an entry point into the AI system or the pipeline that feeds it.
Reaching the model itself — via API, embedded application, or exposed weights — a tactic unique to ATLAS.
Running attacker-controlled logic within the model's environment or a tool it can invoke.
Maintaining a foothold across sessions, retraining cycles, or memory stores.
Avoiding detection by content filters, guardrails, or monitoring tooling.
Mapping the system's architecture, tools, and data sources from the inside.
Harvesting training data, prompts, or model parameters for later use.
Preparing AI-specific attacks such as crafted adversarial examples — the other ATLAS-unique tactic.
Stealing model weights, training data, or sensitive output from the system.
Disrupting, degrading, or manipulating the AI system's behavior or availability.
cat ./nist-ai-rmf.md
The NIST AI Risk Management Framework is voluntary guidance published by NIST (January 2023, with a Generative AI Profile added in July 2024) for building trustworthiness considerations into how an AI system is designed, built, deployed, and evaluated. Instead of prescribing controls, it organizes risk management into four interconnected functions meant to run continuously across a system's lifecycle, not as one-time gates.
Govern
Establish policies, roles, accountability, and oversight for AI risk across the organization.
Map
Understand the system's context, purpose, stakeholders, data, and potential impacts.
Measure
Assess, test, and monitor AI risks and trustworthiness characteristics with real metrics.
Manage
Prioritize, mitigate, respond to, and continuously improve controls based on what Measure finds.