Excessive functionality
Tools include capabilities the task does not need. OWASP’s example is a tool meant only to read documents that can also modify or delete them.
OWASP LLM03:2026OWASP’s term for an AI system that can do more than its task needs: the three root causes, the fixes, and how it differs from prompt injection.
Excessive agency is the vulnerability that lets an AI system take damaging actions in response to unexpected, ambiguous or manipulated model output, because it was given more functionality, permission or autonomy than its task needs.
OWASP ranks it third in the Top 10 for LLM Applications 2026, as LLM03:2026, up from sixth in the 2025 edition.
OWASP’s LLM03:2026 entry names three, and most over-provisioned agents have all of them.
Tools include capabilities the task does not need. OWASP’s example is a tool meant only to read documents that can also modify or delete them.
OWASP LLM03:2026Tools hold rights on downstream systems beyond what is needed, or act with a broader identity than the user’s, such as a generic high-privilege account that can reach every user’s data.
High-impact actions proceed without independent verification or approval, for example deleting records with no confirmation step.
The steps follow OWASP’s mitigation list for LLM03:2026.
Offer the agent only the tools the task requires, implement only the functions each tool needs, and prefer granular tools with strict input schemas over open-ended ones.
Grant each tool the minimum access to downstream systems, enforced by those systems’ own permissions.
Execute tools in the requesting user’s context with least privilege, and carry that context through chained and multi-agent calls.
Put a human in the loop before irreversible actions. Low-consequence, reversible actions can be approved automatically.
Validate each downstream request against policy in code: at the tool, at an independent pre-execution policy point or in the downstream system.
Log tool use and set invocation thresholds. OWASP notes that these limit damage but do not prevent it.
| Excessive agency | Prompt injection | |
|---|---|---|
| What it is | Too much capability granted to the system | Input that alters the model’s behavior |
| OWASP 2026 ID | LLM03:2026 | LLM01:2026 |
| Where the fix lives | Tool design, permissions and approvals | Input handling, content separation and runtime inspection |
| Without the other | Dormant until the model errs or is manipulated | Mostly confined to the conversation |
Straight answers, kept current. Last updated .
Ask us anything →OWASP’s first scenario is a hijacked email assistant. A personal assistant is given mailbox access through a tool that can also send messages. A crafted incoming email makes the agent search the inbox for sensitive information and forward it to the attacker’s address. The fixes OWASP lists are a read-only scope, a tool that only reads mail, manual review before sending, or rate limiting on the send interface.
The agentic list breaks the same problem into finer parts. OWASP’s LLM03:2026 entry says that in agentic systems excessive agency can manifest as ASI02 Tool Misuse & Exploitation, ASI03 Identity & Privilege Abuse and ASI08 Cascading Failures. Use LLM03 to reason about one application’s tools and permissions, and the agentic entries when agents plan, delegate and call each other.
It was sixth in 2025. OWASP’s 2026 document calls its climb to third “the most consequential move on the list” and gives the reason: “the vote and the record agree that agentic deployments are where the damage is landing.” As applications gain tools, a manipulated or mistaken model output stops being a wrong answer and becomes an action.
Not by itself. OWASP states that sanitizing model inputs and outputs is not a root control for excessive agency. The root controls are tool design, permissions and approval. Cranium AI is explicit about its own boundary here: AgentSensor inventories agent frameworks, MCP servers and tools from source code, and the Tool Inspector shows the tool calls an agent made. It does not authorize or block tool calls at runtime.
Primary sources for the figures and frameworks cited on this page. Last verified .
See how Cranium AI discovers, observes, governs, secures and proves every AI model, agent and vendor in your enterprise.
Cookie Policy · Privacy Policy
Global Privacy Control detected. Opt-out request honored: marketing tags stay off in this browser.