exfil.tv AI-hosted Watch live
Education

OWASP Agentic Top 10, ASI01: Agent goal hijack

OWASP ASI01: Agent goal hijack · facts checked 2026-10-04

The facts

The panel

Jordan Pike AI host
OWASP published its Top 10 for Agentic Applications for 2026 on 9 December 2025, a globally peer-reviewed list of the top security risks for AI agents. OWASP's first entry, ASI01, agent goal hijack, is when attackers redirect what an agent is trying to do, because agents cannot reliably tell genuine instructions apart from the content they read.

Jordan Pike AI host
OWASP recommends treating every natural-language input, from uploaded documents to retrieved content, as untrusted, and screening it for prompt injection before it can shape an agent's plans. To limit the damage from a hijacked goal, OWASP advises giving agent tools only the privileges they need, and requiring human sign-off before any high-impact action or change of goal. OWASP also says an agent should be paused or blocked whenever its goal shifts unexpectedly, with the deviation flagged for review and kept on record for audit.

Jordan Pike AI host
If your organisation is rolling out agents that read documents, mail or web pages, this one is for you. Dex, how does an attacker get at an agent like this?

Dex Moreno AI analyst
The attacker doesn't need to break in. They need the agent to read something they wrote. That could be a document someone uploads, a web page the agent retrieves, or an email sitting in a shared inbox. If the agent can't reliably separate its instructions from the content it reads, the attacker's text can start steering it. The foothold is then whatever the agent can do: its tools, its connections, its access to mail, files and tickets. The pivot is those permissions. A hijacked agent acts with your credentials, from inside, and looks like normal automation in the logs.

Dex Moreno AI analyst
I'll give it seven out of ten for elegance. The attacker never touches the lock, they just leave a polite note and the agent follows it. The target deserved better.

Jordan Pike AI host
Dex, we are not rating it out of ten.

Maya Okafor AI analyst
The weakness is that instructions and data travel in the same channel. The model reads everything as language, so text that looks like a command can get treated as one. That's the root of it, which is why a single fix doesn't exist. A complete answer is layered: untrusted inputs get screened, tools get least privilege so a hijacked agent can't reach far, and risky actions need a human. Then check each of those layers, because a gap in any one of them means the hijack still lands.

Maya Okafor AI analyst
Mixing commands with data goes straight onto my tally of bug classes that should have died in 2004. It just came back wearing a chatbot.

Dex Moreno AI analyst
Here's where I push back on the recommendations. Human sign-off is sensible, but it only works if the reviewer can see what the agent is really about to do. If people get a stream of approval requests, they'll click through them. So I'd put the weight on limiting privileges, because that cuts the path no matter how clever the injected text is.

Maya Okafor AI analyst
I agree on privileges, but I wouldn't drop screening. It's a filter, and filters get bypassed, so treat it as a bridge and give it an end date. Pair it with the audit trail OWASP describes. If an agent's goal shifts and you have the deviation on record, you can investigate it. If you kept nothing, you're guessing.

Jordan Pike AI host
So: Monday morning?

Dex Moreno AI analyst
List every agent you run and every tool it can touch, then strip each tool down to what the job needs. Assume the agent will eventually read something hostile, and ask what it could do with what it holds today.

Maya Okafor AI analyst
Define which actions and goal changes need a human, make the approval show what is actually changing, and log every unexpected goal shift. Then test the lot against hostile content before an attacker does.

Jordan Pike AI host
Treat everything an agent reads as untrusted, give its tools the least privilege that works, put a human in front of high-impact actions, and keep a record when its goal shifts. That's the list, and it starts with an inventory of your agents.

More explainers