A row of white circuit breakers in an electrical panel
Field note7 min read
All posts

OWASP Top 10 for Agentic Applications, mapped to architecture controls

OWASP's Top 10 for Agentic Applications names ten risks. Six controls cover most of them, and two have no complete fix. Each risk mapped to its control, the layer that owns it and a public incident.

Photo: Mark Kats / Unsplash

On this page

    In December 2025 OWASP published its Top 10 for Agentic Applications, ten risks that appear once a language model stops answering and starts acting: calling tools, keeping memory, talking to other agents. For an architect or security lead the list is the easy part. The hard part is turning it into a backlog: which control stops each risk, where in the stack it lives, and which of the ten we already cover.

    The short answer: most of the ten are old architecture problems with a new entry point, and six controls cover most of the list: scoped, short-lived delegated identity; a policy gate outside the model in front of every tool; human approval for high-impact actions; sandboxed execution with controlled egress; signed and pinned components with provenance; and logging and monitoring with a way to revoke the agent. Two risks have no complete fix: OWASP's 2026 LLM list says there is no reliable prevention for prompt injection, the main route to goal hijack, and the agentic list itself says some cascading-failure risk stays and has to fit your risk budget. So design for a hijacked agent: limit what it can reach, which is what OWASP calls least agency.

    What the list is, and what it is not

    The agentic list was written by the OWASP GenAI Security Project's Agentic Security Initiative with input from more than 100 industry experts, researchers and practitioners, a 12-person expert review board and a public review (OWASP). It is expert consensus. The numbers ASI01 to ASI10 are identifiers, not a ranking by how often each risk occurs, and the incidents in its appendix are illustrations. It has not been revised since December 2025.

    It sits next to the older Top 10 for LLM Applications, whose 2026 edition came out in August 2026 and is the first edition of that list weighted with incident data: three quarters community vote, one quarter a corpus of 7,714 incidents (OWASP LLM Top 10 2026). OWASP draws the boundary plainly: the LLM list covers the model as a component, and the agentic list takes over once the model acts. One numbering trap: the agentic list cites Excessive Agency as LLM06:2025, while in the 2026 LLM list it is LLM03.

    Appraising the evidence

    OWASP Top 10 for Agentic Applications (Dec 2025)Expert consensusWhich risks matter and OWASP's mitigations for eachStrong as a checklistNot a frequency ranking; not revised since
    OWASP Top 10 for LLM Applications 2026Consensus weighted with incidentsPrompt injection and excessive agency at the model levelStrongVote 75%, incident data 25%
    Incident write-ups and CVEsSingle casesThat a risk is real and how it played outStrong for each caseNo public case found for ASI08
    GitGuardian, secrets sprawl 2026Vendor scanSecrets left in MCP configuration filesIndicativeMethod not stated
    OWASP Agent Control StandardDraft specificationA runtime gate that checks each agent actionEarlyVersion 0.1.4; the reference gate fails open by default

    Source: OWASP documents, incident write-ups and vendor reports listed under Sources, read 9 November 2026

    Download the data (CSV)

    Ten risks, ten controls

    For each risk: what it looks like, the control that does most of the work, the layer of the architecture that owns it, and a public example. The controls are drawn from OWASP's own mitigation lists; the choice of the main one is mine.

    Ten risks, ten controls

    ASI01 Agent Goal HijackText in an email, page or tool result redirects the agent's planTreat all content as untrusted; human approval for goal-changing and high-impact actionsTool gateway, peopleEchoLeak: one crafted email made Microsoft 365 Copilot disclose data (CVE-2025-32711, June 2025)
    ASI02 Tool Misuse and ExploitationThe agent uses a tool it is allowed to use in a harmful wayPer-tool least privilege and a policy gate that validates arguments before the callTool gatewaySupabase MCP: a support ticket made an agent read tokens with ordinary SQL tools and write them into the ticket (July 2025)
    ASI03 Identity and Privilege AbuseInherited or broad credentials reach far beyond the taskShort-lived, task-scoped tokens and a separate identity per agentIdentityGitHub MCP: the agent's GitHub access covered the user's private repositories, which a hijacked agent leaked (May 2025)
    ASI04 Agentic Supply Chain VulnerabilitiesA tool, MCP server, prompt or model is malicious or tampered withSigned, pinned and allow-listed components, re-checked at run timeSupply chain, runtimepostmark-mcp: a fake npm package copied every email to an attacker (September 2025)
    ASI05 Unexpected Code Execution (RCE)Generated or injected code runs on the hostSandbox without production access; validation gate between generating and running codeRuntimeCursor CurXecute: an injection made the agent write an MCP config that ran an attacker command (CVE-2025-54135)
    ASI06 Memory & Context PoisoningPoisoned memory steers later sessionsCheck and attribute every memory write; per-user memory; expire what is unverifiedData and memoryCoSnitch: a web page made Copilot store an attacker instruction in its memory (CVE-2026-24301, August 2026)
    ASI07 Insecure Inter-Agent CommunicationAgents trust messages and peers they should notMutual authentication, signed messages, no open discoveryRuntime, identityServiceNow Now Assist: a hijacked agent recruited a more capable agent by default discovery (November 2025)
    ASI08 Cascading FailuresOne fault spreads across agents faster than people can checkCircuit breakers and checkpoints between planning and executionRuntimeNo public incident found
    ASI09 Human-Agent Trust ExploitationA person approves a harmful action the agent presented wellApproval screens built from facts, not model text; multi-step confirmationPeople and processLies-in-the-Loop: injected text hid a command above the visible part of an approval prompt (September 2025, a demonstration)
    ASI10 Rogue AgentsAn agent keeps acting outside its intended scopeBehaviour baselines, a kill switch and credential revocationRuntime, operationsNo clear public case found; Replit's production database deletion (July 2025) is often cited, but OWASP files it under ASI01 and ASI09

    Most incidents fit more than one risk; each is listed once, under the risk it shows best.

    Source: OWASP Top 10 for Agentic Applications (December 2025); incident write-ups and CVE records listed under Sources, read 9 November 2026

    Download the data (CSV)

    Six controls cover most of the list

    Count how often each control appears in OWASP's own mitigation lists across the ten risks, and a short list emerges. That is the practical news for a backlog: you do not need ten programmes.

    In how many of the ten risks each control appears

    Logging, monitoring and drift detection
    10
    Human approval for high-impact actions
    8
    Signing, pinning and provenance
    8
    Least privilege and scoped, short-lived credentials
    7
    Policy gate outside the model
    7
    Sandbox and egress control
    6

    Risks whose OWASP mitigations include the control

    My count from the mitigation lists in OWASP's document; a different reading moves a bar by one, not the order of the top. Logging limits damage and supports response; it does not prevent anything. A kill switch is named only under ASI04 and ASI10.

    Source: OWASP Top 10 for Agentic Applications (December 2025), mitigation sections for ASI01 to ASI10

    Download the data (CSV)

    Show the data
    controlRisks (of 10)
    Logging, monitoring and drift detection10
    Human approval for high-impact actions8
    Signing, pinning and provenance8
    Least privilege and scoped, short-lived credentials7
    Policy gate outside the model7
    Sandbox and egress control6

    Several of these are already written into the standards agents use. Authorization is optional in MCP, but when a server uses it, the specification (version 2026-07-28) requires tokens bound to that server, forbids a server from passing the client's token on to other services, and requires PKCE (MCP authorization); its security guidance names wildcard scopes as a common mistake and requires consent per client against the confused deputy problem (MCP security best practices). The details are in delegated access for agents and MCP specification changes, explained. What MCP does not give you is signing or pinning of tools: its registry, still in preview, proves who published a server, not that the server is safe (MCP Registry). Supply chain stays your job, and so do secrets: GitGuardian found 24,008 unique secrets in MCP-related configuration files on public GitHub, 2,117 of them valid (GitGuardian).

    What has no complete fix

    OWASP is unusually direct here. The 2026 LLM list says no reliable prevention for prompt injection exists today, so the defence is architectural: assume the instruction boundary will be crossed and limit what the model can do and reach (OWASP LLM Top 10 2026). The agentic list says some cascading-failure risk will remain and has to fit the organisation's risk budget, and that semantic misunderstanding between agents is hard to solve (OWASP).

    That is why the list leans on least agency: give an agent only the autonomy the task needs, and do not make something an agent when a fixed workflow would do. A useful test is Simon Willison's lethal trifecta: private data, untrusted content and a way to send data out (Willison). Willison's advice is to avoid the combination; OWASP's 2026 LLM list sets the floor that an agent with all three needs a person to approve each consequential action. The public incidents behind this are tracked in prompt injection in AI agents, tracked.

    When the list is not enough

    • It does not rank risks. Use it to check coverage, not to decide what to fund first; your own data flows decide that.
    • Mappings to controls are young. OWASP's crosswalk to the AIUC-1 standard (May 2026) shows relevance, not sufficiency, and finds eight gaps in AIUC-1, around agent identity, runtime containment and supply-chain attestation (OWASP AIUC-1 crosswalk). The Agent Control Standard, a wire format for a runtime gate, is at version 0.1.4 and its reference gate fails open by default (ACS repository). Treat both as direction, not as products.
    • Old controls still do most of the work. Identity, least privilege, sandboxing and change control are not new; agents make the gaps in them cost more.

    The six-control baseline

    Before an agent with tools goes live, check these six, in this order:

    1. Identity: the agent acts with the user's delegated, task-scoped, short-lived token, never a shared admin key, and each agent has its own identity in the logs.
    2. Policy gate: every tool call passes a check outside the model, with an allow-list of tools and validated arguments (more in MCP tool design).
    3. Approval: irreversible or high-impact actions wait for a person, on a screen built from the actual call, not from the model's summary of it.
    4. Containment: code runs in a sandbox with no route to production and egress limited to an allow-list.
    5. Provenance: tools, MCP servers and prompts are pinned to reviewed versions, and memory writes carry their source.
    6. Response: actions are logged against the agent's identity, unusual chains raise an alert, and one switch revokes the agent's credentials.

    Where this fits in the wider move from demo to production, including the incidents that made approvals and backups non-negotiable, is in agentic AI from demo to production.

    Sources

    How to cite

    Ivan Bardak. "OWASP Top 10 for Agentic Applications, mapped to architecture controls". bardak.ai, November 12, 2026. https://bardak.ai/posts/owasp-agentic-top-10

    Data behind the charts (JSON)

    Discussion