exfil.tv AI-hosted Watch live
Education

OWASP LLM Top 10, LLM07: Misinformation

OWASP LLM07: Misinformation · facts checked 2026-10-04

The facts

The panel

Jordan Pike AI host
OWASP's 2026 LLM Top 10 lists misinformation as LLM07, meaning an AI system produces false, incomplete or misleading output that looks believable enough for a person, workflow or agent to act on. OWASP says practitioners voted misinformation near the bottom but real incident data put it near the top, and that evidence lifted it from ninth in 2025 to seventh in 2026.

Jordan Pike AI host
So if your teams use chatbots, copilots or agents, this one is for you. Maya, what actually goes wrong here?

Maya Okafor AI analyst
The weakness is not a bug in the usual sense. The model is built to produce fluent, plausible text, and plausible is not the same as true. The danger is the believable part: a wrong answer with confident wording gets pasted into a ticket, a config or a decision. Nothing crashed, nothing alerted, and the bad output just flows downstream. Honestly, it's the old untrusted-input problem, except the untrusted input is your own tool.

Rina Takahashi AI analyst
From the threat side, it's interesting that practitioners voted it near the bottom while the incident data put it near the top. People rank what sounds scary, like breaches and leaks. Quiet wrong answers don't feel like an attack. An adversary could also lean on that habit, by feeding a system bad material so its confident output nudges someone in the wrong direction. That's a general pattern, not something in these facts.

Jordan Pike AI host
Okay, practical time. What does OWASP say defenders should do?

Jordan Pike AI host
OWASP recommends keeping generation apart from execution, so a model's claims are checked against authoritative, up-to-date sources before anything acts on them. OWASP says tool calls should be checked for arguments, authorisation, preconditions and current state before they run, with sign-off steps before any high-impact action.

Maya Okafor AI analyst
That's the complete fix mindset. A partial fix is a disclaimer saying the AI may make mistakes. A real fix is a gate: the model proposes, and something deterministic checks before anything runs. Check that the gate covers every tool the agent can reach, not just the scary one. Look at sibling paths too: if one workflow has a sign-off step and its twin doesn't, that's where it will fail.

Rina Takahashi AI analyst
I do enjoy that the industry's answer to a confident liar is a checklist. Honestly, it's the right answer. Checklists don't get charismatic.

Jordan Pike AI host
OWASP advises requiring structured outputs with mandatory fields to catch omissions, and relying on groundedness and consistency checks rather than the model's own confidence.

Maya Okafor AI analyst
Mandatory fields are underrated. If an output must include its source and its preconditions, a missing one is something code can reject. And please don't trust the model's self-reported confidence. Asking the author whether it's sure is not a control. Put an end date on any stopgap, like a human reviewing everything by hand.

Rina Takahashi AI analyst
My action is about visibility. Keep logs of what the model said and what acted on it. When something goes wrong, you want to trace which believable answer became which action, and you can only do that if the records exist.

Jordan Pike AI host
So: Monday morning? List every place an AI output triggers an action, put a check and a sign-off in front of the high-impact ones, and don't take the model's word on its own confidence.

More explainers