The OWASP Top 10 for Agentic Applications is a community list of the ten main security risks of autonomous AI agents, ASI01 to ASI10, published by the OWASP GenAI Security Project in December 2025. One entry, ASI03 Identity and Privilege Abuse, is about identity directly, and identity controls appear in the mitigations of most of the others.
What the list is, and its status
The document is titled OWASP Top 10 For Agentic Applications 2026. The OWASP GenAI Security Project dates it 9 December 2025, and its Agentic Security Initiative publishes it under the CC BY-SA 4.0 licence. Each entry has an identifier from ASI01 to ASI10, a description, example scenarios and mitigations.
It is a published community framework. It is not a standard of a standards body, it is not law, and nobody can be certified against it. Use it as a shared vocabulary for threat modelling, security questionnaires and internal requirements. It sits beside, not inside, the OWASP Non-Human Identities Top 10, which covers risks to service accounts, secrets and other non-human identities in general.
On 1 September 2026 the same OWASP project said an Agent Control Standard had been donated to it. We cover that separately in our news note; it does not change the Top 10.
The ten entries mapped to identity controls
The table restates each entry briefly and lists the mitigations in the OWASP text that concern identity, credentials, authorisation, approval, revocation or logging. The column mapping each entry to the OWASP NHI Top 10 is our interpretation; OWASP does not publish that mapping.
| Entry | Risk in brief | Identity controls named in the OWASP text | Nearest OWASP NHI Top 10 entry (our interpretation) | Guide on this site |
|---|---|---|---|---|
| ASI01 Agent Goal Hijack | An attacker redirects the goal an agent is pursuing. | Least privilege for agent tools; human approval for high-impact or goal-changing actions; logging and monitoring of agent activity. | NHI5 Overprivileged NHI (privilege sets the blast radius) | Prompt injection as an identity problem |
| ASI02 Tool Misuse and Exploitation | An agent misuses a legitimate tool, for example after prompt injection or unsafe delegation. | Per-tool least-privilege profiles; explicit authentication for each tool invocation; just-in-time, ephemeral credentials; immutable logs of tool invocations. | NHI5 Overprivileged NHI; NHI7 Long-lived secrets | Governing MCP servers |
| ASI03 Identity and Privilege Abuse | Dynamic trust and delegation are exploited to escalate access and bypass controls. | Short-lived, task-scoped tokens; per-action authorization through a central policy engine; human approval for privilege escalation; isolated agent identities; agents managed as non-human identities with scoped credentials, audit trails and lifecycle controls; automated revocation on idle or anomaly. | NHI5 Overprivileged NHI; NHI9 NHI reuse | Delegation chains |
| ASI04 Agentic Supply Chain Vulnerabilities | Third-party agents, tools and related artefacts are compromised or malicious. | Mutual authentication and attestation between agents; no open registration; signed inter-agent messages; an emergency kill switch for specific tools or agent connections. | NHI3 Third-party NHI | Agent revocation and kill paths |
| ASI05 Unexpected Code Execution (RCE) | Code an agent generates or runs is turned into code execution. | Per-session isolation with permission boundaries; never run as root; human approval for elevated runs; logging of generation and execution. | NHI8 Environment isolation | Agent credentials in CI/CD |
| ASI06 Memory and Context Poisoning | Stored context or memory is corrupted to steer later reasoning. | Least-privilege access to memory; segmentation of sessions and contexts; only authenticated, curated sources. | NHI8 Environment isolation | Prompt injection as an identity problem |
| ASI07 Insecure Inter-Agent Communication | Messages between agents are spoofed, intercepted or altered. | Per-agent credentials and mutual authentication; protocol authentication bound to agent identity; signed agent cards verified before accepting discovery or coordination messages. | NHI4 Insecure authentication | Multi-agent authorization and A2A |
| ASI08 Cascading Failures | One fault propagates across connected agents and systems. | Short-lived, task-scoped credentials per agent run; an external policy engine separating planning from execution; tamper-evident logs of inter-agent messages and policy decisions. | NHI5 Overprivileged NHI; NHI7 Long-lived secrets | Incident response for AI agents |
| ASI09 Human-Agent Trust Exploitation | People over-trust a fluent agent and approve what they should refuse. | Multi-step approval before sensitive data access or risky actions; tamper-proof records of user queries and agent actions. | No direct equivalent | Human-in-the-loop patterns |
| ASI10 Rogue Agents | A malicious or compromised agent acts outside its function or authorised scope. | Per-agent cryptographic identity attestation; kill switches and credential revocation; signed audit logs; keys never directly available to agents. | NHI1 Improper offboarding; NHI2 Secret leakage | Agent registries |
Two patterns stand out. First, short-lived, task-scoped credentials and least privilege recur in ASI02, ASI03 and ASI08, so one control answers three entries. Second, tamper-evident logging recurs in ASI02, ASI08, ASI09 and ASI10. An organisation that has neither will find most of the list open, regardless of how good its prompt filtering is.
ASI03 in detail: identity and privilege abuse
ASI03 describes attacks that exploit the dynamic trust and delegation in agents to escalate access and bypass controls. The OWASP text links the risk to agents that lack a distinct, governed identity, which makes true least privilege hard to enforce.
Its mitigations read like a non-human identity programme applied to agents:
- Issue short-lived, narrowly scoped tokens per task.
- Re-verify each privileged step with a centralised policy engine.
- Require human approval for high-privilege or irreversible actions.
- Isolate agent identities and contexts, for example per-session sandboxes with separated permissions and memory.
- Bind OAuth tokens to a signed intent covering subject, audience, purpose and session.
- Treat agents as managed non-human identities with scoped credentials, audit trails and lifecycle controls, including automated revocation on idle or anomaly.
Two of these have no settled standard behind them. Signed intent is a recommendation, not a published protocol, and the NIST NCCoE summary of comments on its agent identity concept paper (29 September 2026) reports no clear consensus on how to implement signed statements of user intent through a broadly accepted standard. For per-call delegation, the ratified building block is OAuth 2.0 Token Exchange (RFC 8693); how to apply it to agents is described in the IETF working group draft draft-ietf-wimse-aims, which is not an RFC. See AI agent authentication and agent identity standards for status.
An identity checklist with acceptance criteria
The OWASP list says what to mitigate, not how to show it is mitigated. The checklist below turns the identity mitigations into tests an internal or external auditor can run. The acceptance criteria are our interpretation and are not part of the OWASP document.
| Control | Entries it addresses | Acceptance criterion (our interpretation) |
|---|---|---|
| Every agent has its own identity, recorded in an inventory with a named owner | ASI03, ASI10 | Each production agent appears in the registry with a named person as owner and the owning team; no agent shares an identity with another agent or with a human. |
| Credentials are short-lived and scoped to the task | ASI02, ASI03, ASI08 | No agent holds a static API key; token lifetimes are documented; a sample of issued tokens shows expiry and scope matching the task. |
| Privileged actions are authorised per call by a policy outside the agent | ASI03, ASI08 | Each privileged tool call in a sample has a matching policy decision record (allow or deny, policy reference, time). |
| High-impact and irreversible actions need human approval | ASI01, ASI02, ASI03, ASI09 | A written list of actions needing approval exists; approval records name the approver and the action; approval is collected outside the agent's own interface. |
| Agents authenticate each other | ASI04, ASI07 | Inter-agent calls without a valid credential are rejected in a test; where A2A is used, agent card signatures are verified. |
| Credentials never enter model context | ASI10 | A review of prompts, tool outputs and memory stores in a sample finds no secrets; secret scanning covers agent memory and logs. |
| A tested kill path exists | ASI04, ASI10 | A drill records the time from revocation to the agent being unable to act, against a target the owner has signed off. |
| Logs are complete and tamper-evident | ASI02, ASI08, ASI09, ASI10 | Log entries for a sample of tool calls show agent identifier, delegating user, tool, decision and timestamp; log integrity protection is documented. |
The approval criterion goes slightly further than the OWASP text. It follows the IETF draft draft-ietf-wimse-aims (section 10.7), which says an agent must not treat local UI confirmation alone as sufficient authorisation. The credentials criterion matches the same draft (section 8), which says the language model must not have access to an agent's credentials.
What the list does not do
- It does not specify protocols. It names controls such as mutual authentication and signed agent cards without fixing a mechanism. For mechanisms and their status, use the standards tracker.
- It does not assign ownership. Owner assignment, access review and retirement are lifecycle questions covered in the AI agent lifecycle.
- It is not evidence of compliance. Mapping a deployment to ASI01 to ASI10 shows coverage of a community risk list. It does not show conformity with the EU AI Act, ISO/IEC 42001 or any other law or standard.
Frequently asked questions
What is the OWASP Top 10 for Agentic Applications?
It is a community list of ten security risks specific to autonomous AI agents, numbered ASI01 to ASI10, published by the OWASP GenAI Security Project on 9 December 2025 under CC BY-SA 4.0. It covers goal hijack, tool misuse, identity and privilege abuse, supply chain, code execution, memory poisoning, inter-agent communication, cascading failures, human trust exploitation and rogue agents.
Which OWASP agentic risk covers agent identity?
ASI03, Identity and Privilege Abuse. It describes attackers exploiting dynamic trust and delegation in agents to escalate access. Its mitigations include short-lived task-scoped tokens, per-action authorization by a central policy engine, human approval for privilege escalation and managing agents as non-human identities. Identity controls also appear under ASI02, ASI04, ASI07 and ASI10.
How does the Agentic Top 10 relate to the OWASP NHI Top 10?
They are separate, complementary OWASP lists. The NHI Top 10 covers risks to non-human identities in general, such as offboarding, secret leakage and overprivilege. The Agentic Top 10 covers threats specific to agents. An agent holds a non-human identity, so deployments usually need both. OWASP does not publish a mapping between them; the mapping on this page is our interpretation.
Can an organisation be certified against the OWASP Agentic Top 10?
No. It is a published community framework, not a standard from a standards body and not law, and no certification scheme exists for it. Use it for threat modelling, questionnaires and internal requirements. Claims of OWASP Agentic Top 10 compliance describe a self-assessment and should come with the evidence behind each entry.
Sources
- OWASP Top 10 for Agentic Applications for 2026, OWASP GenAI Security Project, 9 December 2025
- OWASP Top 10 For Agentic Applications 2026 (PDF), OWASP GenAI Security Project, Agentic Security Initiative, December 2025
- draft-ietf-wimse-aims-00, AI Identity Management System, IETF WIMSE working group, 15 September 2026
- Summary of comments on the concept paper, NIST NCCoE, 29 September 2026
- RFC 8693, OAuth 2.0 Token Exchange, IETF, January 2020
From the publisher
HumanAudit, which publishes this site, is preparing the Agent Governance pack: an agent policy, inventory and owner register, delegated-authority standard, and kill-switch and revocation procedure. Planned for the first week of November 2026; not yet on sale, and no price has been published.