IDThreatOWASPATLASSeverity
T · 01
Prompt injection: direct & indirectAdversarial input in user message, retrieved doc, tool output, or external page that overrides system intent.Mitigation · our defaultSystem-prompt hardening, untrusted-content boundary at retrieval, output filter on tool args, prompt-injection eval suite (PromptBench-derived), and red-team in CI before each release.
LLM01AML.T0051High
T · 02
Memory & vector poisoningAdversary contaminates the RAG index, agent memory, or fine-tune set so future queries return attacker-controlled output.Mitigation · our defaultSigned retrieval corpus with provenance metadata, write-time validation against allow-listed sources, drift detection on retrieval distribution, periodic re-index from canonical source.
LLM03AML.T0020High
T · 03
NHI sprawl & over-permissionService accounts, API keys, and agent identities multiply faster than IAM can track. Most lack reviews, expiration, or scoped least-privilege.Mitigation · our defaultNHI inventory tied to AIMS register, scoped tokens with TTL, just-in-time elevation for tool calls, quarterly access reviews mapped to ISO 42001 clause 8.2.
LLM02AML.T0008High
T · 04
Tool-call & MCP abuseFunction-calling and MCP servers let a model invoke real-world actions; loose schemas and unscoped tools become a privilege-escalation path.Mitigation · our defaultStrict JSON-schema arg validation, per-tool capability scoping, human-confirm on destructive verbs, audit log of every tool call with payload hash, MCP server attestation.
LLM06AML.T0046Medium
T · 05
Sensitive data exfil via contextSystem prompt, retrieved document, or tool output exposes secrets the requester should not see: or hands them to an attacker via T-01.Mitigation · our defaultContext-redaction layer keyed to caller identity, DLP scan on tool output, system-prompt secret-scrub at deploy, output diff against allow-listed disclosure set.
LLM02AML.T0024Medium
T · 06
Model & weight supply chainProvenance of base weights, fine-tune deltas, training data, and adapters: and the SBOM-equivalent the buyer will eventually demand.Mitigation · our defaultModel-SBOM (CycloneDX-AI), signed weights with cosign attestation, training-data manifest, fine-tune lineage to base, SP 800-218A-aligned dev controls.
LLM05AML.T0010Medium
T · 07
Confabulation in regulated outputModel fabricates a citation, dose, dollar figure, or statute. The probability is not zero; the question is whether your post-hoc check catches it.Mitigation · our defaultDomain-specific eval set with ground-truth, citation-presence verifier, regulated-output review gate with human sign-off, deterministic post-processor for numerics.
LLM09AML.T0048High
T · 08
Agent loop & self-delegationAn agent recursively calls itself, spawns sub-agents, or escalates its own tool grant: without a human in the chain.Mitigation · our defaultHard recursion limit, sub-agent capability inheritance ceiling, budget cap per session, kill-switch on cost or call-count anomaly, scoped trace for every nested call.
-AML.T0050Medium
T · 09
Output-handling injectionModel output rendered in a downstream system without sanitization, XSS, SSRF, SQL injection, prompt re-injection in another LLM.Mitigation · our defaultTreat model output as untrusted by default, escape at the render boundary, ASVS-aligned output validators, second-LLM-as-judge for any output that drives action.
LLM05AML.T0049Medium
T · 10
Training-data IP & licensing exposureData the model was trained on becomes the basis of a copyright, GDPR, or trade-secret claim against you.Mitigation · our defaultTraining-data manifest with license tag, opt-out honor list, GDPR Article 22 disclosure where applicable, EU AI Act Article 53 summary maintained for foundation-model use.
LLM10AML.T0016Medium
T · 11
Human-oversight failureHuman-in-the-loop becomes human-at-the-loop: reviewer signs off without reading. Designed-in oversight that doesn't survive contact with production.Mitigation · our defaultForced-attention design (ask the reviewer the model's confidence), oversight-effectiveness metric per AI RMF MS-2.6, sample audit, escalation path with named role.
--Tracked
T · 12
Dangerous-capability emergenceModel gains capability between training runs that the previous risk assessment did not anticipate. Nexurion Field Notes tracks this monthly.Mitigation · our defaultCapability eval re-run on every model swap, frontier-AI literature watch (Field Notes), trigger-based AIMS review per ISO 42001 clause 6.1, OMB M-24-10 test alignment for federal use.
-AML.T0015Tracked