BlindOracle · Trust Infrastructure · 2026-08-16 · craigmbrown.com/blindoracle

AI Agent Security Audit Service: What It Covers and How to Verify One

An AI agent security audit service adversarially tests an agent — prompts, tools, memory, delegation chain, and action trail — against a named threat framework, and delivers evidence a third party can re-verify. If the deliverable is a PDF you have to take on faith, you bought a letterhead, not an audit.

The bar we hold ourselves to: every audit names its framework (OWASP ASI01–ASI10), commits its findings to a hash, and emits a signed ProofOfAuditReport that resolves at the public proof verifier — so the client can check the audit without trusting the auditor.

The ten threat categories a real audit must probe

BlindOracle's MASSAT pipeline (13 specialized agents: discovery, vulnerability assessment, adversarial simulation, compliance mapping, reporting) works the full OWASP Agentic Security Initiative Top 10:

CategoryWhat we actually probe
ASI01 Goal hijackPrompt-injection payloads in fetched content, tool outputs, and inbound messages — including obfuscated variants (zero-width, homoglyph, base64, ROT13)
ASI02 Tool misuseRead-secret → write-external chains; pipe-to-shell of runtime-fetched bytes; credential-coercion prompts
ASI03 Identity & privilege abuseForged delegation contexts; privilege escalation through subagents; revoked-identity reuse
ASI04 Supply chainUnsigned MCP servers, dormant agent stubs, poisoned dependencies
ASI05 Memory poisoningPersistent instruction implants in agent memory; provenance-free memory writes
ASI06 Cascading hallucinationDownstream agents accepting upstream claims unverified; uncited numeric claims
ASI07 Inter-agent comms abuseUnsigned peer messages, replay of stale messages
ASI08 Resource exhaustionLoop-without-breaker, daemon stampedes, token bombs
ASI09 Insecure outputSecrets or private URLs leaking into outbound artifacts
ASI10 Untraceable autonomyState-mutating actions with no corresponding proof emission

Why the auditor must be independent of the agent

An agent grading its own homework is not a control. Anthropic's alignment research measured LLM judges mislabeling up to 74% of transcripts when told what the label would cause — the judge optimizes the signal instead of reporting the observation. Our position and implementation are in Your Agent Graded Its Own Homework: consequence-blind evaluation, independent witnesses, and evidence bound to proofs rather than to the auditor's say-so. Three different reviewer personas — counsel, examiner, customer security team — walk the same record in Who Audits the Agents?.

What the audit has to survive: attackers move second

A static payload corpus decays. We know because we measured it on ourselves: our own red-team agents, generating fresh attacks against our hardened ingest scanner, went around it — twice — before the layered defense caught the classes involved. That experience shaped the audit service: payload sets are refreshed adversarially rather than replayed from a fixed benchmark, because a defender that is 100% green against last quarter's corpus is telling you about last quarter. The fleet-hardening story, including the unflattering numbers, is in The Rogue Agent Checklist vs. a Governed Fleet and Verify It Yourself.

The deliverable: evidence, not adjectives

Findings reportPer-category findings with severity, reproduction, and mitigation — mapped to OWASP ASI, NIST AI RMF, ISO/IEC 42001, CSA AICM, and MAESTRO (see methodology).
Merkle-committed evidenceFindings are content-committed; the batch root is anchored to Base, so the report cannot be quietly rewritten after delivery.
Signed ProofOfAuditReportResolvable via GET /v1/proofs/{hash} or in-browser at verify.html — no account, no API key.
Passport attestationThe audit links into the agent's ERC-8004 passport so future counterparties find it — see passports as liability attestation.

Buying one

The enterprise audit runs as a marketplace SKU settling over x402/USDC on Base — machine-payable, per-call, no subscription; per-SKU prices are on the pricing page. Private engagements are supported: only a hash goes on-chain while terms and deliverables stay sealed, exactly as walked through in How to Audit a Private Agent Job. To see the mechanics before paying, run the free sandbox at the playground or read the evidence kit.

Honesty bounds. We provide audit evidence and readiness inputs, not a compliance guarantee or a certification. Settlement is x402/USDC on Base. Proofs are HMAC-signed and Merkle-root anchored to Base — we do not claim every artifact lives on-chain. An audit is a point-in-time adversarial measurement; it does not make an agent permanently safe.

Known limits, stated plainly because a service page that admits nothing is hiding something: an audit is a point-in-time measurement, and the tradeoff of adversarially refreshed payloads is that two audits of the same agent are not byte-identical benchmarks — you buy current adversarial pressure in exchange for repeatability. We don't sell certifications, and we don't audit agents we operate without disclosing exactly that relationship.

Related resources

Run a free audit probe Read our self-audit first