Mapping agent controls to the OWASP Top 10 for LLM apps

The OWASP Top 10 for Large Language Model Applications, 2026 release, lists ten risks from prompt injection (LLM01) to improper output handling (LLM10). Most agent controls address more than one of them. Tool scoping, a policy check outside the model, human approval for high-impact actions, spend caps and logs cover most rows of the table below.
This post maps the LLM list only, using the 2026 release on GitHub. OWASP’s agentic risk list is a separate document and is not covered here. The mappings are vendor-neutral. The last section says where a product feature is involved.
The ten risks and the controls that address them
The table names the controls each risk page lists as mitigations, in the page’s own terms. Read the full page before you write a control into a policy, because the table is a summary.
| Risk | Agent controls that address it |
|---|---|
| LLM01:2026 Prompt Injection | Keep credentials and state-change capability in application code, behind a deterministic policy check. Require approval for privileged actions. Limit any agent that combines untrusted input, sensitive data and external action. |
| LLM02:2026 Sensitive Information Disclosure | Enforce document and chunk authorization inside the retrieval query. Send external providers only the fields a task needs. Keep secrets out of prompts. Restrict and scrub logs. |
| LLM03:2026 Excessive Agency | Minimize tools and their functions. Avoid open-ended tools. Minimize tool permissions. Run tools in the user’s context. Require user approval. Mediate every call with a policy decision point. Monitor and rate-limit tool use. |
| LLM04:2026 Supply Chain | Vet suppliers. Keep a signed inventory (SBOM, extended to models and datasets). Pin and verify third-party tool packages. |
| LLM05:2026 Data and Model Poisoning | Filter and score retrieved content. Sandbox interaction with unverified data and tools. Put human oversight on automated retraining. Version datasets so you can roll back. |
| LLM06:2026 Unbounded Consumption | Set rate limits and quotas on tokens and cost. Set hard spending caps per key, user and team. Set step, recursion and per-run cost limits for agents. Restrict what the app can reach over the network. |
| LLM07:2026 Misinformation | Verify claims before acting. Validate tool-call arguments, authorization and current state. Use runtime verification for high-impact actions. |
| LLM08:2026 Hidden Context Exposure | Keep secrets out of system prompts. Enforce authorization outside the LLM. |
| LLM09:2026 Vector and Embedding Weaknesses | Enforce tenant scope inside the index query. Separate indexes by trust tier. Keep immutable logs of retrieval activity. |
| LLM10:2026 Improper Output Handling | Treat model output as untrusted input to backend code. Use parameterized queries. Encode output for the sink it reaches. Disable automatic rendering of images and links in clients by default. |
Five controls that cover most of the table
Five controls show up in more than one row. Most of them are decisions made outside the model, so they still hold when the model has been tricked.
Scope each tool before you grant it
The Excessive Agency page starts with removal. Take away any tool the agent does not need. Then cut each remaining tool to one function. A mail summary tool needs read access, not send or delete. Replace open-ended tools such as a shell or a URL fetch with narrow tools that validate input against a strict schema.
Finally, give each tool the least permission on the downstream system. The page’s example is a product recommender that only needs read access to a products table, so it gets no insert, update or delete rights. The same page says to run tools in the user’s context and to carry that context through chained agent calls. A shared service identity should not do the work. A narrow tool limits what an injected instruction can do even when the injection succeeds.
Decide outside the model
The Excessive Agency page’s complete mediation rule says authorization belongs in logic, not in an LLM’s judgement about whether an action is allowed. Each request to a downstream system is checked by the tool, by a policy decision point between the tool and the system, or by the system itself.
LLM08 makes the same point about hidden context. Privilege separation and authorization bounds checks must not be delegated to the model, whether through the system prompt or anywhere else. LLM01 asks for a deterministic policy engine that re-validates intent and arguments at execution time. The Excessive Agency page also describes a graduated policy of audit, warn, block or escalate, chosen by how reversible the action is.
Require approval for high-impact actions
Human approval appears in LLM01, LLM03 and LLM07. LLM03 asks for a human to approve high-impact actions, either in the tool itself or in a downstream system. LLM01 asks for approval before any privileged, irreversible or externally visible action, and for the reviewer to see the exact action rendered, not a summary of it.
LLM01 also warns that approval fatigue degrades reviewer judgement at volume, so approve the actions that matter rather than every call. Its Rule of Two sets the floor. An agent with untrusted input, sensitive data and state change or external communication at once needs per-action approval. An agent with only two of the three needs a documented residual-risk assessment.
Cap spend, steps and rate
LLM06 separates alerts from limits. An alert on a budget tells you after the spend has happened, and fast workloads can outrun it. The page asks for non-overridable ceilings per API key, user, team and cloud account that halt inference when they are reached.
For agents, LLM06 adds step limits, recursion depth limits, time limits and per-run cost ceilings, with state hashing to catch loops. LLM03 adds circuit breakers that halt, rate-limit or escalate when a threshold is crossed. It also suggests context-aware thresholds, based on the cumulative value of a tool parameter rather than only the call count. LLM06 asks you to baseline normal tool behaviour first, so the thresholds come from your own traffic.
Record every tool call with its context
LLM03 says to log and monitor tool activity, in the tools and in the downstream systems, so that undesirable actions can be found and answered. LLM09 asks for immutable logs of retrieval activity, with tenant scope, query, returned IDs and similarity scores. LLM10 asks for monitoring that flags unusual patterns in model output.
Logs answer the first question an incident responder asks: what did the agent try, and with which arguments. We would also keep the blocked calls, since they show what an injected instruction was after. LLM02 adds a caution. Logs and traces can hold the sensitive data you are trying to protect, so restrict and scrub them before they reach a shared observability tool.
What the list does not promise
Two fine-print points matter more than any single control. They change how you should test and how you should read a vendor’s claim.
Prompt injection has no reliable fix yet
The LLM01 page says no reliable prevention mechanism exists today, because the model makes no architectural distinction between instructions and data. Its defense is architectural: design the system on the assumption that the instruction boundary will eventually be bypassed, and constrain what the model can do and what its output can reach.
The page cites Nasr et al. (2025), who found static attack success near zero while adaptive attack success exceeded 90% for most of 12 recent defenses. OWASP uses that result to reject static-only attack-success claims. A test suite that the defense’s authors already know is not evidence of safety.
Pinning is not the same as safe
LLM01 says to pin, sign and verify every MCP server and third-party tool package, and to audit tool descriptions for hidden instructions. It then warns that pinning does not stop a payload shipped in the pinned version. It also does not stop tool-description poisoning that leaves the version number unchanged. Review the description text of every pinned tool as well as its version.
Checklist for an agent review
- List every tool the agent can call, with the function and the downstream permission behind each.
- Flag any agent that combines untrusted input, sensitive data and external action. Require per-action human approval for that combination. Where a residual-risk assessment is also needed, write it separately for each configuration. It does not replace approval.
- Confirm that each allow decision is made by code outside the model.
- Set per-user and per-run ceilings on cost, steps and recursion, and confirm they halt the run rather than only alerting.
- Pin each MCP server and tool package, and review its description text when it changes.
- Log tool calls, blocked calls and retrieval scope, and scrub the logs before they leave the system.
- Test against an attacker who has read the defense, and keep a static prompt set as the baseline.
Enforcing this with AIControls
AIControls is Nirmata’s product for governing AI agents in Kubernetes. Four of its capabilities map to rows above: routing Claude Code through the gateway for the CLI and VS Code, per-user tokens, per-user and per-session cost and budget limits, and PII that is held for approval or redacted or blocked.
The table in this post is a test for any product, AIControls included. Map each control in your deployment to a row, and mark the rows that nothing covers. For the allowlist side of this, read tool-level authorization for AI agents. For the input side, read prompt injection defense for agents. The AIControls product page describes the platform, and the governing AI agents hub covers the wider pillar.
Frequently asked questions
Does the OWASP LLM Top 10 cover agentic AI?
The 2026 LLM Top 10 covers LLM applications. Its Excessive Agency entry, LLM03, is the one that deals with tools and autonomy. OWASP's agentic risk list is a separate document, and this post does not map it. Check that list on its own before you claim full coverage of an agent deployment.
Is prompt injection solved in the 2026 list?
No. The LLM01 page says no reliable prevention mechanism exists today, because models do not architecturally separate instructions from data. OWASP's advice is architectural: assume the instruction boundary will be bypassed, and limit what the model can do and reach, so a successful injection does bounded damage. Test against adaptive attackers, and do not stop at static prompt sets.
Which control covers the most rows of the OWASP LLM Top 10?
Mediating each tool call with a decision made outside the model. The 2026 text lists it in the mitigations for LLM01, LLM03, LLM07 and LLM08. LLM07 asks for tool calls to be validated before they run, but the decision outside the model is stated in LLM01 and LLM03. LLM01 asks for a deterministic policy engine that re-validates intent and arguments at execution time, so the model may propose an action but does not decide whether it runs.
Should blocked tool calls go in the audit log?
We would keep them. LLM03 says to log and monitor tool use so undesirable actions can be found, and a blocked call shows what an agent or an injected instruction tried to do. LLM02 adds a caution: logs can hold the sensitive data they record, so restrict and scrub them before they reach a shared observability tool.