⭐ Star on GitHub

Summary: Controls organised by control rather than by threat, each showing which frameworks in this vault call for it. Multi-framework agreement is the ranking signal.

Sources: synthesis across raw/MITRE-ATLAS/, raw/Laws/, raw/Agentic/, raw/Supply-chain-attacks/, raw/Data-poisoning/, raw/AI-incident-response/, raw/OWASP-Top-10-for-LLMs/, raw/Model-theft/

Last updated: 2026-07-30


Every source here organises defences by the threat they answer. That is the wrong axis for anyone building — you implement a control once and it covers many threats. This page inverts it.

Framework keys: ATLAS = MITRE ATLAS mitigations · CoP = EU Code of Practice Appendix 4 · ASI = OWASP Agentic Security Initiative playbooks · T10 = OWASP Top 10 for LLMs · SC = supply chain sources (Google, OpenSSF, Plexal/Oxford) · DP = data poisoning sources (CMU, health care review) · IR = incident response sources (CoSAI, Microsoft) · MAE = CSA MAESTRO · OBS = agent observability implementation (Realm) · ENG = agent reliability engineering (Galileo)

MAE is not an independent vote. Ken Huang authored MAESTRO and contributes to the OWASP Top 10, so MAE agreeing with ASI or T10 does not raise a control’s tier. Its controls are also stated at framework level and mapped to no individual threat — ~14 controls for ~60 threats. Where MAE appears below it is for coverage no one else offers, not for agreement. See the crosswalk.

One finding here changes how to read the whole page. MAESTRO Layer 6 models attacks on AI systems used as security tooling — evasion, poisoning their operational data, extracting the detection model to learn how to bypass it. Every control below that is itself AI-based — guardrails, anomaly detection, alignment critics, trust scoring, the entire model-checks-model patterninherits the full threat model of the thing it protects. No other source in the corpus says this.

Tier 1 — Four or more frameworks agree

These are the controls to implement first. Agreement across independently-built frameworks is the strongest evidence the corpus offers.

ControlFrameworksNotes
Telemetry and audit loggingATLAS CoP ASI IR OBS MAE + AI Act Art 12The precondition for everything else. OBS is the corpus’s only field-level schema; MAE is the only source that treats the log as a target. See telemetry-and-logging
Least privilege / access controlATLAS CoP ASI T10 DPApplies to model, agent, tools, and training data alike
Sandboxing and execution isolationATLAS CoP ASI SCCoP requires it against model self-exfiltration; ASI requires destroy-after-execution
Provenance / cryptographic signingSC DP ASI ATLASSigning extends to prompt templates and tool definitions in ASI
Rate limitingATLAS CoP ASI T10Token-aware, not request-count. Also the anti-extraction control
Red teamingCoP ASI SC DPCoP distinguishes infrastructure red teaming from model adversarial testing — different exercises, different teams

Tier 2 — Two or three frameworks

ControlFrameworksNotes
Input and output validationATLAS T10 SCLLM05 implemented; output treated as untrusted input
Guardrails / content filteringATLAS T10 ASIProbabilistic control against an adversary with unlimited attempts
Human approval for sensitive operationsASI SC T10Degrades under load — see T10
MFA / strong authenticationCoP ASIASI extends to mutual agent-to-agent authentication
Data validation before trainingDP CoP SCCoP: “inspect training data for tampering or sabotage”
Anomaly / drift detectionDP ASI IRHardest part is distinguishing attack from natural drift
Session isolationASI T10Prevents cross-user memory contamination
Model alignment and refusal trainingATLAS CoPCoP Commitment 5; ATLAS AML.M00*
SBOM / dependency inventorySC ASIASI adds AI SBOM / AIBOM / Agent SBOM
Encryption at rest and in transitCoP SCCoP specifies 256-bit and TPM key storage
Staged / phased deploymentDP IR CoPCoP allows phasing “to correspond with model capability advancement”
Personnel vetting and insider programmesCoP SCCoP scopes insider threat to include “by or through models”

Tier 3 — Single framework, still worth having

Distinctive controls that only one source names. Several are among the most concrete in the corpus.

ControlSourceWhy it matters
Attested TEEs for parameters in useCoP 4.2Closes the memory-dump path. Materially higher bar than most production ML infra meets
Secure registry of every device holding weightsCoP 4.2You cannot protect copies you have not enumerated
Output rate limiting on interfacesCoP 4.3Anti-exfiltration; also constrains gradient-based and best-of-N attacks
Chain of custody signing the data–label pairingDPAnswers the attacker who “modif[ies] some labels, leaving the images themselves unaltered”
MEDLEY ensemble disagreement monitoringDPPreserve disagreement rather than collapsing to consensus; monitor patterns, not cases
Spotlighting / context isolationSCStructurally mark untrusted context. Its own mechanism is an attack surface
User Alignment Critic isolated from poisoned contextSC (Google)Isolation is the load-bearing property
Pin container images by immutable digestSC (OpenSSF)Verify with cosign/notation, enforce via admission controller
Limit the agent’s ability to print linksASI T15Narrow, cheap, directly defeats link-rendering exfiltration
Destroy or reset the sandbox after each tool executionASI PB3Blocks persistence and lateral movement between invocations
Just-in-time tool access with immediate revocationASI PB3Grant on demand only
Topology segmentation of multi-agent systemsASI PB6Shrinks the graph an attacker traverses
Decision approval discrepancy trackingASI PB6Catches orchestration hijacking, which looks clean to every individual agent
Knowledge lineage trackingASI PB2Does not prevent cascades; the only control that lets you find the origin and roll back
AI gateway as security control planeSCOne chokepoint to instrument — precondition for telemetry
Responder wellbeing / rotation protocolsIRExplicitly costs capacity; “a deliberate investment, not a free addition”
Procurement security requirementsSCAlready functioning as a control: 28% of orgs had changed vendors over trust issues
Correlation IDs spanning agent and tool actionsOBSMakes the task the investigable unit, not the log line. The cheapest item on this page and a precondition for attributing any tool call to the principal that caused it
Principal identity on every tool invocationOBSprincipal_user_id. Without it the confused deputy is not merely exploitable but unauditable
Cap the agent’s tool countENGQuantified: 46 tools failed where 19 succeeded, independent of context-window space. Least privilege on toolsets now has a reliability argument attached — the version an engineering team will actually act on
Per-agent context isolationENG SCSub-agents with independent windows and a coordinator seeing only summaries. Blast-radius containment for context poisoning, and the strongest empirical case in the corpus: +90% performance at 15× the tokens. Does not work where subtasks interact (e.g. coding)
Cap context length for guardrail-critical agentsENGInstruction adherence degrades past ~100k tokens; at threshold models “ignore instructions entirely”. No source frames this as a control — it follows from the measurement
Approved-registry enforcement for MCP servers and agentsMAE OBSMAE Layer 7 names registry compromise and malicious agent discovery; OBS supplies source_mcp_server as the field that detects drift from the approved list
Treat AI-based controls as in-scope for the threat modelMAELayer 6. Not a control so much as a scoping rule, and the corpus’s only statement of it

Controls named but not solved

Prescribed somewhere in the corpus while the underlying capability does not reliably exist. Treat as research directions, not roadmap items.

”Control”SourceStatus
Cryptographic attestation of agent/model integrityASI T13Source states plainly these mechanisms “do not yet exist” for LLMs
Probabilistic truth-checking of new knowledgeASI PB2No method given for establishing truth
Agent trust scoringASI PB6No scoring function specified
Deception detection / truthfulness verification modelsASI T7Threat itself conceded to be “at an early stage”
Behavioural consistency analysisASI T7, T9Requires a model of honest behaviour that nobody has
Assured service verificationSC”Lack of metrics and standards for assured service provision”
Formal verification of agent behaviour and goal alignmentMAEStated in a single clause with no method, no tooling and no tractability caveat. The most ambitious “control” in the corpus
Immutable audit trailsASI OBSUniversally prescribed; the corpus’s one implementation asserts immutability without implementing it — nothing signs, chains or write-protects the log. See telemetry-and-logging
Correction of a poisoned context by instructionNot prescribed anywhere, but widely assumed. Measured to fail: corrections and the original error coexist with comparable weight, and “the agent trusts its own context more than external corrections”

The model-checks-model pattern

Appearing in ASI (T6, T7, T9, T15), SC (Google’s alignment critic), and DP (MEDLEY ensembles). It is the default answer wherever rules cannot catch the threat, and it has one genuine design rule attached, from Google via StackHawk:

the critic must be isolated from the potentially poisoned context.

A checker sharing the context inherits the compromise. In a multi-agent setting the checker also becomes part of the T13 surface it was deployed to police. MEDLEY’s version is the most defensible because its diversity is architectural — “models with different architectures, training algorithms, and data origins are unlikely to share identical vulnerabilities.”

MAESTRO makes the objection structural rather than contextual. Its Layer 6 is the whole pattern’s threat model written out: a security agent can be evaded adversarially, have its operational data poisoned so it misidentifies threats, have its model extracted so an attacker learns the bypass, or be taken over and used to disable other controls. Isolation from the poisoned context — Google’s rule — answers only the second of those four.

Two consequences for this page. First, every ASI control resting on a model checking a model is weaker than its tier suggests. Second, context isolation earns its place above on both readings: it is the one intervention that limits what a compromised checker can see and what a poisoned context can reach.