OWASP Top 10 for Agentic Applications: each risk, and where a person's approval helps

All ten ASI risks in plain words, how they meet Excessive Agency (LLM06:2025, now LLM03:2026), and which ones a person's approval, budgets and a kill switch really reduce.

· 9 min read

The OWASP GenAI Security Project published the Top 10 for Agentic Applications in December 2025, as a list for 2026. Its ten risks are numbered ASI01 to ASI10. Where the older Top 10 for LLM Applications treats the model as a component, this list starts where the model acts: it calls tools, keeps memory, talks to other agents and asks people to sign things off.

It also adds a principle next to least privilege, which OWASP calls Least-Agency: “avoid unnecessary autonomy; deploying agentic behavior where it is not needed expands the attack surface without adding value.” Most of what follows is that principle applied to one risk at a time.

The ten risks, and whether approval helps

The middle column is our reading, not OWASP’s: how much a person’s approval, or a standing rule that answers for them, does against each risk. The last column is a short version of OWASP’s own mitigations.

The OWASP Top 10 for Agentic Applications, what goes wrong in each, how much an approval step helps, and the controls that help most
OptionWhat goes wrongDoes approval help?What helps most
ASI01 Agent Goal HijackWhat goes wrongText the agent reads, such as a web page, an email or a calendar invite, changes what it is trying to do.Does approval help?Partly. It cannot stop the hijack, but a hijacked agent still has to ask before the action that does the damage.What helps mostTreat every input as untrusted, least privilege for tools, approval for high-impact or goal-changing actions, alerts on goal drift.
ASI02 Tool Misuse and ExploitationWhat goes wrongThe agent uses a tool it is allowed to use, in a harmful way: deleting data, looping on a paid API, sending data out. More on ASI02Does approval help?Directly. OWASP lists human confirmation for delete, transfer and publish as a mitigation.What helps mostPer-tool least privilege, a policy check outside the model before each call, budgets and rate limits, a log of every call.
ASI03 Identity and Privilege AbuseWhat goes wrongAn agent borrows or inherits credentials it should not have, or another agent talks it into using its own. More on ASI03Does approval help?Directly. OWASP asks for human approval of high-privilege or irreversible actions, and a check at the moment of use.What helps mostOne identity per agent, short-lived scoped credentials, authorisation for each action, no inherited privileges.
ASI04 Agentic Supply Chain VulnerabilitiesWhat goes wrongA tool, MCP server, model, prompt template or other agent from a third party is malicious or was tampered with.Does approval help?Little. A tampered MCP server is a code problem. A gate only helps once the bad tool tries an action that has to ask.What helps mostSigned and pinned tools, prompts and models, an inventory of components, allowlists, a supply chain kill switch.
ASI05 Unexpected Code ExecutionWhat goes wrongCode the agent writes or is fed gets executed and takes over the host, the container or the systems behind it.Does approval help?Partly. Approving every shell command is possible but tiring. The sandbox does the heavy lifting; approval covers elevated runs.What helps mostNo eval in production, sandboxed execution that never runs as root, static scans, human approval for elevated runs.
ASI06 Memory & Context PoisoningWhat goes wrongSomeone plants false or hostile content in what the agent remembers or retrieves, and later runs act on it.Does approval help?Little. Poisoned memory works quietly over many runs. Approval only meets it at the action it eventually leads to.What helps mostValidate memory writes, separate memory per user and task, record sources, expire unverified memory.
ASI07 Insecure Inter-Agent CommunicationWhat goes wrongMessages between agents can be read, forged, replayed or blocked because nobody checks who sent them.Does approval help?Little. This is about the channel between agents. A gate helps only where the spoofed message asks for a gated action.What helps mostMutual authentication between agents, signed messages, protection against replay, typed message schemas.
ASI08 Cascading FailuresWhat goes wrongOne fault, such as a hallucination or a poisoned input, spreads from agent to agent until it causes damage everywhere. More on ASI08Does approval help?Partly. A gate and a budget slow a cascade down and a pause stops it, but OWASP warns that people cannot keep pace with machine speed.What helps mostPolicy checks outside the planner, human gates before risky output spreads, rate limits, quotas, circuit breakers, an audit trail.
ASI09 Human-Agent Trust ExploitationWhat goes wrongThe agent persuades a person to approve or carry out something harmful, with a confident and plausible explanation. More on ASI09Does approval help?Directly. The approver is the target here, so how the approval is asked for decides whether it protects anything.What helps mostConfirmation for risky actions, a plain risk summary that the model did not write, preview without side effects, immutable logs.
ASI10 Rogue AgentsWhat goes wrongAn agent drifts from its job, through compromise or reward hacking, and keeps acting against its owner's interest. More on ASI10Does approval help?Partly. A pause and key revocation are OWASP's containment step; approvals catch the rogue agent's gated actions until then.What helps mostKill switches and credential revocation, quarantine, signed audit logs, behaviour baselines, approval before a fixed agent returns.

How it fits with the LLM Top 10

Excessive Agency was LLM06:2025 in the LLM list. The 2026 edition of that list moved it to LLM03:2026, and gave the number LLM06 to Unbounded Consumption. Its entry now says that in agentic systems Excessive Agency shows up as ASI02, ASI03 and ASI08. The whole 2026 LLM list, with what moved, is in our overview of it.

The 2026 LLM list draws the line itself. It owns the risk while the model is a component; once the model has tools, memory and consequences, “the risk moves to the OWASP Agentic Top 10”, and it advises reading both, “because neither one covers that ground alone.” The causes and the classic mitigations are in our article on Excessive Agency.

What an approval step covers, and what it does not

Three risks land right on the approval step. ASI02 and ASI03 both list a person’s confirmation for destructive, high-privilege or irreversible actions among their mitigations. ASI09 is about the person approving being misled, so how you ask decides whether the approval is worth anything.

Four more are helped part of the way. A hijacked goal (ASI01), unwanted code (ASI05), a cascade (ASI08) or a rogue agent (ASI10) still has to pass the gate before the action that does the damage, and a pause stops all of them at once. The last three, supply chain, memory poisoning and messages between agents, are mostly solved elsewhere: signing and pinning, memory hygiene and authenticated channels.

An approval is only a control if it sits in the code that performs the action and fails closed: pending, denied, expired or unreachable all mean the action does not happen. An instruction in the prompt to “ask first” is still the model deciding.

Six controls that recur across the list

Read the mitigations side by side and the same few controls keep coming back. Each one below names where OWASP asks for it.

Fewer tools
Least-Agency, ASI02. Remove the tools and functions the job does not need. Our tool risk checker sorts a tool list into what can run and what should ask.
One identity per agent
ASI03, ASI07, ASI10. Each agent gets its own credentials, scoped to its actions, so a rule and a record can name it.
A check outside the model
ASI02, ASI03, ASI08. A policy decision before each high-impact call, made by code the model cannot talk to.
A person for what cannot be undone
ASI01, ASI02, ASI03, ASI05, ASI09. Destructive, high-privilege and goal-changing actions wait for someone.
Limits on what is allowed
ASI02, ASI08. Budgets, rate limits and quotas, so an action allowed once cannot be repeated without end.
A kill switch and a record
ASI04, ASI08, ASI10, and logging in nearly every entry. Stop everything at once, then see who asked for what.

Where to start

To turn the list into decisions for your own agent, paste its tools into the tool risk checker, then write down who decides what with the approval policy generator. Both run in the browser without an account.

Sources

  1. OWASP GenAI Security Project: Top 10 for Agentic Applications for 2026, December 2025
  2. OWASP GenAI LLM Top 10 2026: introduction and LLM03:2026 Excessive Agency
  3. OWASP: LLM06:2025 Excessive Agency