⭐ Star on GitHub

Summary: Four control sets in this vault — ATLAS mitigations, Code of Practice Appendix 4, OWASP ASI playbooks, OWASP Top 10 prevention guidance — what each is for, what each protects, and why no two are substitutes.

Sources: raw/MITRE-ATLAS/AML.M00*.md; raw/Laws/Overview of the Code of Practice...md; raw/Agentic/Agentic-AI-Threats-and-Mitigations-1.1.pdf; raw/OWASP-Top-10-for-LLMs/

Last updated: 2026-07-29


The vault records four attack taxonomies. It also holds four control sets, and they divide the problem just as cleanly — but nobody says so, because each is published attached to its own taxonomy.

Side by side

ATLAS mitigationsCoP Appendix 4ASI playbooksTop 10 prevention
Count35 mitigations5 control families6 playbooks, ~90 controls10 entries
GranularityPer techniquePer control familyPer threat groupPer risk
ProtectsThe deployed systemUnreleased weightsAgent behaviour and seamsThe LLM application
BindingNoVoluntary (GPAI systemic-risk)NoNo
Structuremitigates pairs with per-pairing textPrescriptive requirementsProactive / reactive / detectiveProse per risk
Agentic coverageSubstantialAlmost noneTotalPartial (LLM06)
Weight protectionThin256-bit, TPM, attested TEEs, copy registryNoneNone
ImplementabilityHigh — concrete, per-techniqueHigh — specifies numbersMixed — a third name capabilities that don’t existModerate
Known coverage gap56% of techniques have zero mitigationsNothing on prompt injection or agentic containmentNo secure-design methodGranularity too coarse for agents

What each is actually for

ATLAS mitigations — the only technique-level control catalogue in the corpus. Each mitigates relationship carries its own applicability text, so the same control reads differently against different techniques. Use it when you have a specific technique and need to know what answers it. The 35 split 16 Technical–AI / 12 Technical–Cyber / 7 Policy, which carries its own message: roughly a third of defending AI systems is ordinary security engineering.

Code of Practice Appendix 4 — a hardening baseline for the parameter file. Read as a whole it defends against weight theft, not misuse of a deployed model. It is the most prescriptive set here and the only one with numbers. Use it if you hold unreleased weights that matter; it is nearly silent otherwise.

ASI playbooks — the only set that addresses systems that act. Organised proactive/reactive/detective and scoped by a decision path that prunes against a specific architecture. Use it for anything agentic. Sort its controls before adopting — see the implementability sort.

Top 10 prevention guidance — the shared vocabulary. Too coarse to build from directly, and LLM06 Excessive Agency is doing the work of three separate agentic threats. Use it to name risks for people who need names, then descend to one of the other three.

Why none is a substitute

ATLAS and CoP are close to complementary. Technique-level coverage for the deployed system plus a hardening baseline for the weights. Neither alone is sufficient — and the pairing was already noted in ATLAS mitigations before the agentic material was ingested.

ASI fills the gap both leave. CoP Appendix 4 “has nothing on prompt injection, tool poisoning or agentic containment beyond ‘sandboxing and code execution isolation’.” ATLAS has substantial agentic coverage but at technique granularity, with no architectural scoping. ASI provides the architecture, the decision path, and the seam threats — inter-agent protocols, delegation chains, human-boundary exploitation — that neither of the others models.

None of the four covers secure design. All are runtime or organisational. See defense-in-depth.

Read the coverage gaps as gaps, not coverage

The single most important caution across all four:

97 of 173 ATLAS techniques (56%) have zero mapped mitigations.

Using any of these as a checklist produces a false picture. An unmitigated technique is not one you have handled — it is one where the framework has no published answer. The same reading applies to the seven ASI threats with no Top 10 counterpart, and to CoP Appendix 4’s silence on the agentic surface: absence of a control is a finding, not a pass.

Which to reach for

SituationStart with
You hold unreleased weightsCoP Appendix 4
You are shipping an LLM featureT10, then SC application defences
You are shipping agentsASI, Playbook 3 first
You have a specific technique to answerATLAS
You need to brief non-specialistsT10
You are training or fine-tuningDP chain of custody + CoP 4.4

A fifth set, named but not held

MAESTRO (Cloud Security Alliance) — a layered extension of STRIDE for agentic AI, characterised in the ASI source as offering “a detailed lens to identify Agentic Threats through the use of architectural layers,” covering both agentic and traditional ML/application threats. OWASP declined to adopt it on cognitive-load and scope grounds while recommending practitioners evaluate it. It is the closest thing to a secure-design method any source here names, and no primary source for it is in raw/.