Learn
OWASP Top 10 for Agentic Applications mapped to controls.
ASI01 to ASI10, one row each: the risk, what helps, the MAQPNA control and where it falls short.
Updated 31 October 2026.
What is the OWASP Top 10 for Agentic Applications?
The OWASP Top 10 for Agentic Applications is a list of the ten most critical security risks of AI agents: systems that plan, call tools, keep memory and act with delegated credentials. The OWASP GenAI Security Project published the 2026 edition on 9 December 2025, with items numbered ASI01 to ASI10. It is a voluntary risk catalogue, not a standard you certify against.
It sits next to the OWASP Top 10 for LLM Applications 2025 and does not replace it. The LLM list covers the model and the application around it: prompt injection (LLM01), excessive agency (LLM06), unbounded consumption (LLM10) and so on. The agentic list covers what happens once the model can act. Most agent reviews need both. The Agentic Security Initiative page links the full document.
How does each ASI risk map to a runtime control?
Names are as OWASP publishes them. "Mitigates" means the root cause lies elsewhere, usually in the model. The same mapping, in short, is on the trust centre.
| ASI id · risk | What helps | MAQPNA control | Coverage | Docs |
|---|---|---|---|---|
| ASI01 · Agent Goal Hijack | Treat retrieved content as untrusted; least privilege that holds whatever the model decides; gate exfiltration after untrusted input | Taint tracking holds or denies sinks (push, send, post) after a session reads untrusted content; injection guards on tool output; per-tool policy | Mitigates: contains the effect, the model can still be fooled | Taint |
| ASI02 · Tool Misuse | Least agency: allow only the tools and arguments a task needs; rate limits; approval for destructive calls | Policy per tool and argument (typed conditions, Cedar, Rego), default deny, rate limits, budgets, approvals; policy test, replay and eval before rollout | Enforced | Policies |
| ASI03 · Identity & Privilege Abuse | A distinct, short-lived identity per agent; act on behalf of a named person; no shared keys | Signed session token per session with an on-behalf-of claim, per-call scope checks, narrowing-only delegation, upstream credentials kept in the gateway or a vault, optional SVID mTLS | Enforced | Identity |
| ASI04 · Agentic Supply Chain Vulnerabilities | Pin and verify tools, images and servers; keep an inventory of what agents actually use | Tool pinning (a changed tool definition is hidden and denied), registry allow-lists, digest-pinned images, cosign-verified releases, a CycloneDX AI-BOM | Enforced for tools and images; mitigates for third-party models | Tool pinning |
| ASI05 · Unexpected Code Execution | Run generated code in a sandbox with no route out except a checked one | A sandbox per session (gVisor, microVM or confidential VM), non-root, read-only root, no service-account token, default-deny network with the gateway as the only exit | Contained | Sandboxes |
| ASI06 · Memory & Context Poisoning | Scope memory per agent and user; scan what is written; expire and erase | Governed memory stores scoped per agent and user, DLP on writes, retention, erasure with a signed certificate; ephemeral sessions | Mitigates | DLP and taint |
| ASI07 · Insecure Inter-Agent Communication | Authenticate peers; allow-list who may call whom; keep the delegation chain | Governed A2A: signed agent cards, peer allow-lists, nested delegation chain with depth limits, policy and audit on both hops | Enforced inside one installation; cross-organisation clearing is planned | MCP and A2A |
| ASI08 · Cascading Failures | Limit blast radius; hard spending and rate limits; fail closed | Sandbox and network limits, budgets with hard stops, rate limits, session TTLs, fail-closed defaults, model fallback with circuit breakers | Mitigates | Budgets |
| ASI09 · Human-Agent Trust Exploitation | Show approvers what will really happen; separate duties; confirm with the affected user | Approvals bound to the exact call, four-eyes quorum, no self-approval, redacted argument previews, confirmation by the user on their own device (CIBA) | Enforced for held calls | Approvals |
| ASI10 · Rogue Agents | A way to stop an agent at once, and a record to prove what it did | Kill switch across the installation within seconds, session timeline, posture checks, alerts, hash-chained audit ledger | Stop and prove: enforced. Behavioural anomaly scoring: not built in | Kill switch |
What does MAQPNA not cover?
A runtime limits and records what a fooled model can do. Some parts of the list need other controls.
| Gap | Related ASI | Why | What helps instead |
|---|---|---|---|
| The model being persuaded | ASI01, ASI06 | No runtime can stop a model being fooled; MAQPNA contains the effect | Model-side defences, evaluation and red-teaming |
| Behavioural anomaly detection | ASI10 | Not built in | A SIEM or UEBA reading the ledger stream |
| Weights and training data of third-party models | ASI04 | MAQPNA governs the route to a model, not what is inside it | Supplier due diligence, model cards, your own evaluation |
| Wrong but allowed answers | ASI09 | An allowed call with a misleading output is a model-quality problem | Human review of outputs, grounding, user training |
| Agents outside the gateway | All | Controls apply only to traffic that goes through it | Default-deny egress so the gateway is the only way out |
Where MAQPNA fits
How do you show an auditor the OWASP controls work?
A mapping on a web page is a claim. Evidence is a record. MAQPNA builds an evidence pack for the owasp-asi framework from the tamper-evident audit ledger and the configuration in force: one system, one period, signed, with CSV workpapers of approvals, denials by reason, DLP blocks and revocations. Add the AI-BOM and the pack cites it under ASI04 and ASI06.
Each control is marked evidenced, partial or "not covered by MAQPNA (customer process)". A pack supports a review; it does not certify anything, and its mapping stays marked draft until counsel and audit partners review it. Anyone can verify a pack offline. Generating one needs a licence with the evidence-packs feature or the Sovereign edition.
For the controls behind the rows, read prompt injection in agent tool calls (ASI01), what an MCP gateway is (ASI02 to ASI04), human-in-the-loop approval (ASI09) and the kill switch (ASI10). The full threat model is in the security model. To see how it fits a wider programme, read AI agent governance, the EU AI Act for AI agents and ISO 42001 and the NIST AI RMF, or browse solutions by sector.
FAQ
Common questions
What are ASI01 to ASI10?
The ten items of the OWASP Top 10 for Agentic Applications (2026): Agent Goal Hijack, Tool Misuse, Identity & Privilege Abuse, Agentic Supply Chain Vulnerabilities, Unexpected Code Execution, Memory & Context Poisoning, Insecure Inter-Agent Communication, Cascading Failures, Human-Agent Trust Exploitation and Rogue Agents.
How is it different from the OWASP Top 10 for LLM Applications?
The LLM list (2025) covers models and LLM applications, such as prompt injection and excessive agency. The agentic list covers agents that act with tools, memory, credentials and other agents. It extends the LLM list rather than replacing it.
Does MAQPNA make an agent OWASP compliant?
No. The list is a voluntary risk catalogue, and no product makes you compliant with it. MAQPNA supplies runtime controls for each item and evidence that they ran, and it says where it only mitigates or does not cover a risk.
Can a runtime stop agent goal hijack?
It cannot stop the model being fooled. It can stop a fooled model doing damage: taint rules hold or deny exfiltration after untrusted input, and policy limits every tool call whatever the model decided.
Put a control behind every risk.
Run an agent through the gateway on your laptop today, then generate an OWASP evidence pack from the ledger.
curl -fsSL https://maqpna.com/install.sh | shirm https://maqpna.com/install.ps1 | iex