AI Is Now In The Attack Path
Copilots, chatbots and agents connect to your data and tools; ATLAS catalogues how they are being attacked in practice so defences are aimed at real techniques.
MITRE ATLAS (Adversarial Threat Landscape for AI Systems) is a free, public knowledge base of the tactics, techniques and procedures attackers use against machine-learning models, LLM applications and AI agents. Modelled on MITRE ATT&CK, it organises 16 tactics from Reconnaissance to Impact, documents each technique with real case studies, and pairs them with mitigations, so defenders can threat-model, test and monitor AI systems in a shared vocabulary.
ATLAS mirrors the ATT&CK data model, so teams already using ATT&CK can extend their threat modelling and detections to AI without learning a new structure. Select a part to see what it holds and how XcellHost uses it.
The sixteen adversary goals, from Reconnaissance and Resource Development through AI Model Access, AI Attack Staging and Execution to Exfiltration and Impact. Most mirror ATT&CK; AI Model Access and AI Attack Staging are unique to attacks on AI.
How each goal is achieved: prompt injection, poisoning training data, discovering model ontology, evading detection, extracting models through inference APIs, jailbreaks and more. Each technique has a description, examples and links to mitigations and case studies.
Finer-grained variants under a technique — for example direct versus indirect prompt injection, or different ways of publishing poisoned artefacts. They let detections and tests be specific without losing the parent technique's context.
Defensive measures mapped to the techniques they counter — guardrails for generative AI, input restoration, model hardening, access control on models and data, AI honeypots and telemetry logging — so a threat model turns into a control plan.
Documented real-world attacks and red-team exercises against deployed AI — poisoned recommendations, model extraction, agent hijacking, infrastructure compromise through AI services — each mapped to the tactics and techniques used, so the threats are evidence, not theory.
ATLAS links to ATT&CK Enterprise, and MITRE publishes a combined STIX bundle and an ATLAS Navigator layer, so AI techniques sit alongside conventional ones in the same threat matrices, detection coverage maps and intelligence feeds.
Copilots, chatbots and agents connect to your data and tools; ATLAS catalogues how they are being attacked in practice so defences are aimed at real techniques.
Because ATLAS follows the ATT&CK model, existing threat-informed defence, detection engineering and purple-team practice extend to AI systems without a parallel programme.
RBI's FREE-AI report, MeitY's AI governance guidelines and the DPDP Act all expect AI risks to be identified and managed; ATLAS gives the threat side of that evidence.
Boards respond to documented incidents more than abstract warnings; ATLAS case studies show poisoning, extraction and agent hijacking happening to real organisations.
ATLAS has no legal standing in India and no regulator names it, but Indian AI guidance is converging on risk-based governance, and ATLAS is the most complete public catalogue of what AI attacks actually look like, which makes it the natural threat reference for Indian builders and buyers.
| Aspect | MITRE ATLAS | MITRE ATT&CK |
|---|---|---|
| Focus | Attacks on AI and ML systems | Attacks on enterprise, mobile and ICS |
| Unique tactics | AI Model Access, AI Attack Staging | None AI-specific |
| Evidence base | Case studies, incl. red-team exercises | Observed threat-group activity |
| Use together? | Yes — shared data model and STIX bundle | Baseline for everything around the model |
One team for assessment, remediation, 24×7 operations and audit evidence — run from Mumbai by a company certified to ISO/IEC 27001 and ISO/IEC 20000-1.
Talk to our security team →Inventory AI systems and agree which ATLAS tactics and techniques are in play.
Weeks 1–2Build per-system threat models and rank techniques by exposure and evidence.
Weeks 2–4Red-team the top techniques and implement mapped mitigations with owners.
Months 2–4Tag SOC coverage with ATLAS, review gaps quarterly and refresh as releases land.
OngoingXcellHost is certified to ISO/IEC 27001 and ISO/IEC 20000-1 and has protected Indian businesses since 1999 — so your MITRE ATLAS work is run by a team that lives these controls every day.
Adversary tactics, techniques and procedures
Security Operations & ThreatThe ten most critical risks in LLM-powered applications
AI Security & GovernanceComprehensive guidance for securing AI systems and data
AI Security & GovernanceDefensive countermeasure techniques, mapped to attacks
Security Operations & ThreatManaging the risks of AI systems across their life cycle
AI Security & GovernanceSearch the complete master list across 12 domains.
Framework explorer