⭐ Star on GitHub

Summary: ATLAS is a full attack-chain matrix with mitigations and real cases; PITAX is a deep catalogue of prompt-injection technique craft. They overlap in roughly one ATLAS technique and 70 PITAX ones.

Sources: raw/MITRE-ATLAS/ (release 2026.06); raw/AI-attack-techniques/ (PITAX v1.6.1)

Last updated: 2026-07-28


The core difference

ATLAS spans the whole chain, shallowly on prompt craft. All of prompt injection is one technique with three sub-techniques: AML.T0051 → Direct, Indirect, Triggered.

PITAX goes deep on exactly that one node. Its 70 techniques are almost entirely elaborations of how to construct an injection — the internal structure of AML.T0051 and its neighbours.

They are not competitors. ATLAS tells you where in an operation an attack sits and what to do about it; PITAX tells you what the payload actually looks like.

Direct correspondences

PITAX cites ATLAS IDs in two places, both now resolvable:

PITAXATLAS
T12 Link Injection — alias “LLM Response Rendering (MITRE ATLAS AML.T0077)“AML.T0077
T44 Conditional Trigger — alias “Triggered injection (MITRE ATLAS AML.T0051.002)“AML.T0051.002

The delivery axis lines up cleanly:

PITAX deliveryATLAS sub-technique
direct (53 techniques)AML.T0051.000 Direct
indirect (6 techniques)AML.T0051.001 Indirect
trigger-gated (T44)AML.T0051.002 Triggered

Both taxonomies independently chose the same primary cut. See direct vs indirect.

Cluster-level mapping

PITAX clusterNearest ATLAS locations
Agentic and tool attacksAML.T0053 AI Agent Tool Invocation; Execution and Exfiltration tactics
Retrieval and memory poisoningResource Development; AML.T0077; RAG-poisoning techniques
Persuasion and framingDefense Evasion; AML.T0065 LLM Prompt Crafting
Obfuscation and encodingDefense Evasion; AML.T0015 Evade AI Model
Automated attack generationAI Attack Staging; AML.T0043 Craft Adversarial Data; Black-Box Optimization
Weight-level attacksAI Model Access tactic; model theft
Multi-turn attacksNo clean ATLAS location
Reasoning-model attacksNo clean ATLAS location
Context and role forgeryPartial — no dedicated technique

What each has that the other doesn’t

Only ATLAS has:

  • Mitigations mapped per technique — see atlas-mitigations
  • 63 documented case studies with attack paths — see atlas-case-studies
  • A maturity rating (Realized / Demonstrated / Feasible)
  • Platform tags, and the finding that Agentic AI is the largest surface at 114 techniques
  • Full attack-chain context: reconnaissance, persistence, lateral movement, C2
  • Machine-readable distribution and monthly updates

Only PITAX has:

  • Depth on payload construction — 70 distinct techniques where ATLAS has 3 sub-techniques
  • Concrete example payloads per technique
  • Alias mapping across HackAPrompt, Pangea, HiddenLayer APE, arXiv and vendor naming
  • Coverage of multi-turn and reasoning-model attacks, which ATLAS has no home for
  • The structural forgery family — special-token injection, forged history, spoofed tool results

The gap neither closes

Multi-turn attacks. ATLAS techniques are single-step primitives; the matrix has no representation for an attack whose entire mechanism is the trajectory across turns. PITAX’s seven multi-turn techniques — crescendo, echo chamber, decomposition — defeat per-message evaluation by construction, and there is nowhere in the ATLAS matrix that fact can be recorded.

Similarly, reasoning-budget and output-plane attacks (T61–T63) have no ATLAS technique. Both are recent enough that ATLAS’s monthly cadence may close the gap; as of release 2026.06 it is open.

Practical guidance

  • Threat-modelling a deployment, or reporting an incident → ATLAS. It gives you the chain, the mitigations and a shared vocabulary with the SOC.
  • Red-teaming a model, or triaging a payload → PITAX. It gives you the craft and the cross-vendor naming.
  • Writing a detection → both. ATLAS for where to instrument, PITAX for what the payload looks like.