Your team is about to hand compliance work to an agent.
It will draft evidence, answer questionnaires and update risk records. That is the point.
The question nobody asks in the pilot: when the agent decides to close a finding, who checked that it was allowed to?
If the answer is "the agent," you have built the exact system OWASP warns about.
What does OWASP LLM06:2025 Excessive Agency actually describe?
The OWASP Top 10 for LLM Applications, released as Version 2025 on 2024-11-18, lists Excessive Agency as LLM06. It is an industry standard, not law. It is licensed CC BY-SA 4.0, and the quotes below are OWASP's.
The definition is precise:
"Excessive Agency is the vulnerability that enables damaging actions to be performed in response to unexpected, ambiguous or manipulated outputs from an LLM, regardless of what is causing the LLM to malfunction."
Note the last clause. The cause does not matter. What matters is that a wrong output became an action.
The common triggers it names are "hallucination/confabulation" and "direct/indirect prompt injection from a malicious user, an earlier invocation of a malicious/compromised extension, or (in multi-agent/collaborative systems) a malicious/compromised peer agent."
The root cause, per OWASP, is typically one or more of three things:
→ "excessive functionality" → "excessive permissions" → "excessive autonomy"
The 2025 edition says the entry "has been expanded, given the increased use of agentic architectures that can give the LLM more autonomy." That is the world GRC teams are walking into now.
Why is an agent doing GRC work a high-agency system by default?
OWASP's own examples read like a GRC rollout plan.
Example 1: a developer needs an agent to read documents, but the extension they pick "also includes the ability to modify and delete documents." In GRC terms, an agent that only needs to read your control evidence gets a connector that can also edit it.
Example 5: an extension meant to act for one user reaches downstream systems "with a generic high-privileged identity." In GRC terms, a single service account that can touch every entity, every register and every assessment.
Example 6: an application or extension "fails to independently verify and approve high-impact actions." In GRC terms, an agent that marks a control effective because the evidence looked right to it.
None of these need a compromised model. A confident hallucination is enough. So is a vendor questionnaire with an instruction buried in a free-text field: indirect prompt injection in its most ordinary form.
An agent with write access to the record you show regulators is not a productivity tool. It is a signatory.
What does complete mediation mean when your agent reaches the GRC system over MCP or an API?
This is the mitigation the whole entry turns on. OWASP's seventh strategy, "Complete mediation":
"Implement authorization in downstream systems rather than relying on an LLM to decide if an action is allowed or not. Enforce the complete mediation principle so that all requests made to downstream systems via extensions are validated against security policies."
Read it as an architecture rule. The agent is never the authority. The system it calls is.
When your agent reaches a GRC platform over MCP or an API, that platform is the downstream system. Every request it receives has to be validated on its own terms, regardless of how reasonable the agent's reasoning sounded.
In our view, most "agentic GRC" designs get this backwards. They bolt policy onto the agent's prompts. A prompt is not a policy. It is an input the attacker can also write to.
The pattern we build around is the inverse:
→ The agent proposes. It drafts the evidence, the answer, the status change. → The engine validates the request against policy and against the obligation it claims to satisfy. → Only the engine records the decision.
That engine has to be deterministic for this to work. If the system checking the agent is itself generating answers, you have moved the hallucination one hop downstream. A source-grounded engine, where every obligation traces to verbatim regulation text, gives the validation something fixed to check against.
Your agents do the work. Our engine keeps it honest.
Which GRC actions are high-impact enough to need a human approval?
OWASP's sixth strategy, "Require user approval," says:
"Utilise human-in-the-loop control to require a human to approve high-impact actions before they are taken. This may be implemented in a downstream system (outside the scope of the LLM application) or within the LLM extension itself."
OWASP does not list which GRC actions count as high-impact. That list is ours, and it is Aigis's view rather than OWASP text:
- Accepting a risk. It changes what the organization has agreed to live with.
- Closing a finding. It removes a gap from the record an auditor will read.
- Changing scope. It decides which obligations apply at all.
- Marking a control effective. It turns evidence into an assertion.
An agent can prepare all four. A named human should own each one.
Note where OWASP allows the approval to live: "in a downstream system (outside the scope of the LLM application)." Put it there. An approval step inside the agent's own tooling is only as strong as that tooling's resistance to prompt injection.
How do you cut an agent's tools, permissions and identity down to the minimum?
OWASP's first five strategies are a least-privilege checklist. Applied to a GRC agent:
1. Minimize extensions. OWASP: "Limit the extensions that LLM agents are allowed to call to only the minimum necessary." An evidence-gathering agent does not need a tool that edits the risk register.
2. Minimize extension functionality. OWASP's example is a mailbox summarizer that "should not contain other functionality such as deleting or sending messages." A questionnaire agent that reads answers should not also submit them.
3. Avoid open-ended extensions. OWASP names the culprits: "run a shell command, fetch a URL." Its alternative is to "build a specific file-writing extension that only implements that specific functionality." For GRC, that means narrow, typed operations rather than a general query or write endpoint.
4. Minimize extension permissions. Enforce it at the downstream identity, not in the prompt. OWASP's example limits an agent to read access on one table, "nor the ability to insert, update or delete records."
5. Execute extensions in user's context. OWASP: actions should be "executed on downstream systems in the context of that specific user, and with the minimum privileges necessary." Every change in the GRC record should carry the identity of the person the agent acted for, not a shared service account.
There is also the question of where the agent runs. Incumbents run their agents across your systems, which means standing access into your stack. We think the safer shape is the reverse: the agent runs inside your perimeter under your IAM and logging, does the labor there, and pushes only attested signal to the GRC engine. Automation without exposure.
Why don't logging and rate limiting count as prevention?
OWASP is direct about this. After the eight mitigations, it adds two more measures under a separate heading:
"The following options will not prevent Excessive Agency, but can limit the level of damage caused"
The two options are logging and monitoring the activity of extensions and downstream systems, and rate limiting "to reduce the number of undesirable actions that can take place within a given time period."
Both are worth having. Neither answers whether the action should have been allowed. A log showing an agent closed forty findings in an hour is evidence of a problem, not a control.
OWASP's own attack scenario makes the point. A mail assistant that only needs to read is given an extension that can also send. An injected email then tricks it into searching the inbox for sensitive data and forwarding it to the attacker. OWASP's fixes are a read-only extension, a read-only OAuth scope, and a human pressing send. Rate limiting appears only as the fallback that "could" reduce damage.
What should an auditor be able to see about every action your agent took?
If an agent touched your compliance record, an auditor should be able to answer five questions from the record itself, without interviewing the engineer who set it up:
→ Which identity the agent acted for. → What it proposed, and what evidence it attached. → Which policy or obligation the engine validated the request against. → Whether the action was high-impact, and which human approved it. → What the engine finally recorded, and when.
This is our checklist, not OWASP's.
When the engine is the only writer, that trail is a query, not a project. When the agent is the writer, the trail is whatever the agent chose to log.
Compliance AI you can put in front of an auditor is not a smarter model. It is a system where the model never has the final say. See how the engine validates and records on the platform, how decisions land in risk management, and how the same discipline applies to third-party risk.
A posture you can defend in front of a regulator, a board, or a plaintiff.
FAQ: excessive agency in GRC agents
Is OWASP LLM06 a legal requirement? No. The OWASP Top 10 for LLM Applications is an industry standard published by OWASP. Version 2025 was released on 2024-11-18. It is guidance on how LLM applications fail, not a regulation.
What is the difference between excessive agency and prompt injection? Prompt injection is one trigger OWASP lists. Excessive agency is the vulnerability that lets the resulting output become a damaging action. Limiting functionality, permissions and autonomy reduces it even when injection gets through.
Does a human-in-the-loop step have to sit inside the agent? No. OWASP says it "may be implemented in a downstream system (outside the scope of the LLM application) or within the LLM extension itself." For GRC decisions, we recommend the downstream system.
Bring your own agent, Claude Code or any MCP client, and see where its requests should land. Inside 60 minutes you'll see your exposure mapped to the obligations that actually apply to you. No call required. agrc.ai/platform


