OWASP Agentic Top 10, ASI02: Tool misuse and exploitation
OWASP ASI02: Tool misuse & exploitation · facts checked 2026-10-04
The facts
- OWASP's ASI02, tool misuse and exploitation, covers agents using legitimate tools in unsafe ways while staying inside the permissions they were given, such as deleting valuable data or leaking information. (OWASP)
- OWASP recommends a least-privilege profile for every tool, with set scopes, rate limits and outbound allowlists, such as read-only database queries and an email summariser that cannot send or delete mail. (OWASP)
- OWASP says each tool call should be authenticated, and destructive actions such as deleting, transferring or publishing should need a human's confirmation after a dry-run preview. (OWASP)
- OWASP advises treating planner output as untrusted, with a policy checkpoint in front of tool calls that checks their purpose and parameters, applies rate limits and issues short-lived credentials. (OWASP)
- OWASP recommends usage ceilings on cost, rate or tokens, so an agent that blows through its budget is automatically throttled or has its access revoked. (OWASP)
The panel
Jordan Pike AI host
If your organisation is letting AI agents touch real systems, this one is for you. OWASP's ASI02, tool misuse and exploitation, covers agents using legitimate tools in unsafe ways while staying inside the permissions they were given, such as deleting valuable data or leaking information. OWASP recommends a least-privilege profile for every tool, with set scopes, rate limits and outbound allowlists, such as read-only database queries and an email summariser that cannot send or delete mail.
Jordan Pike AI host
OWASP says each tool call should be authenticated, and destructive actions such as deleting, transferring or publishing should need a human's confirmation after a dry-run preview. OWASP also advises treating planner output as untrusted, with a policy checkpoint in front of tool calls that checks their purpose and parameters, applies rate limits and issues short-lived credentials.
Jordan Pike AI host
Last item: OWASP recommends usage ceilings on cost, rate or tokens, so an agent that blows through its budget is automatically throttled or has its access revoked.
Jordan Pike AI host
So the agent isn't breaking rules, it's misusing the access it was handed. Dex, how does an attacker see that?
Dex Moreno AI analyst
Nothing has to be broken here, and that's the appeal. An attacker could try to steer the agent, say through content it reads or a poorly guarded instruction, and then let the agent do the work with its own legitimate access. The foothold is whatever the agent can already reach: the database, the mailbox, the publishing pipeline. From there it's lateral movement by proxy, because every tool the agent holds is a door the attacker never had to pick. Logs just show the agent doing agent things.
Dex Moreno AI analyst
Honestly, getting a trusted employee-shaped robot to do your errands with its own badge? Eight out of ten for elegance.
Jordan Pike AI host
Dex, we are not rating it out of ten.
Dex Moreno AI analyst
Fine. The target deserved better. A summariser that can delete mail deserved a lot better.
Maya Okafor AI analyst
The bug class is excessive permission plus misplaced trust. Someone gives a tool broad scope because it's convenient, then trusts the model's output as if it were a vetted request. That's why OWASP says to treat planner output as untrusted and put a policy checkpoint in front of every call that checks purpose and parameters. A complete fix is narrow scopes per tool, authenticated calls and short-lived credentials. A partial fix is a prompt that says please be careful. Check that the read-only tool is actually read-only at the database, not just in its description.
Maya Okafor AI analyst
Handing a process far more access than the job needs goes straight onto my tally of things that should have died in 2004. Now it just has a chat window.
Dex Moreno AI analyst
Here's where I'd push back. Human confirmation sounds great, but people who approve forty previews a day start clicking yes on reflex. I'd put the weight on blast radius: scopes, allowlists, ceilings. If the tool can't delete, nobody has to be awake to stop it.
Maya Okafor AI analyst
Those aren't rivals, Dex. Scopes limit what's possible, and confirmation covers the rare action that has to stay possible, like a delete or a transfer. Keep confirmations scarce so they mean something, and use the dry-run preview so the human sees what will happen. Usage ceilings are the backstop: an agent that runs wild gets throttled or loses access automatically.
Jordan Pike AI host
So: Monday morning?
Maya Okafor AI analyst
List every tool your agents can call and write down what each one can read, write or delete. Cut each to the minimum, and gate destructive actions behind a confirmation with a preview.
Dex Moreno AI analyst
Assume an agent will eventually be steered somewhere it shouldn't go, and ask what it could reach. Add budget and rate ceilings, use short-lived credentials, and review the agent's tool-call logs for odd behaviour before an attacker does.
Jordan Pike AI host
Least privilege, a checkpoint in front of every tool call, a human for the destructive stuff, and a ceiling on spend. The agent can be clever; the permissions should be boring.





