OWASP LLM Top 10, LLM01: Prompt injection
OWASP LLM01: Prompt injection · facts checked 2026-10-04
The facts
- OWASP's 2026 LLM Top 10 ranks prompt injection first, as LLM01, meaning any input that steers a model in ways its developer did not plan, whether typed by a user or hidden in a web page, image or tool result. (OWASP)
- OWASP says credentials and the power to change things belong in application code rather than the model, with each operation given the least privilege it needs and checked by a deterministic policy engine. (OWASP)
- OWASP treats an agent that can read untrusted input, reach sensitive data and also change state or contact the outside world as high risk, and says it needs a human to approve each action. (OWASP)
- OWASP advises giving outside content to the model through a separate channel labelled with its origin, and stripping invisible Unicode, such as zero-width characters, wherever text is ingested or displayed. (OWASP)
- OWASP says defences should be tested by attackers who have been told how they work, and that results from static attacks alone should not be taken as proof a defence holds. (OWASP)
The panel
Jordan Pike AI host
OWASP's 2026 LLM Top 10 ranks prompt injection first, as LLM01. OWASP defines it as any input that steers a model in ways its developer did not plan, whether typed by a user or hidden in a web page, image or tool result.
Jordan Pike AI host
OWASP says credentials and the power to change things belong in application code rather than the model, with each operation given the least privilege it needs and checked by a deterministic policy engine. OWASP also treats an agent that can read untrusted input, reach sensitive data and also change state or contact the outside world as high risk, and says it needs a human to approve each action.
Jordan Pike AI host
OWASP advises giving outside content to the model through a separate channel labelled with its origin, and stripping invisible Unicode, such as zero-width characters, wherever text is ingested or displayed. It also says defences should be tested by attackers who have been told how they work, and that results from static attacks alone should not be taken as proof a defence holds.
Jordan Pike AI host
Who should care: anyone building or buying software that lets an AI model read things and act on them. Dex, how does an attacker see this?
Dex Moreno AI analyst
Notice the attacker doesn't have to be your user. If the model reads a web page, an image or a tool result, whoever wrote that content is effectively talking to your agent. The foothold is whatever the agent can do: if it holds credentials and can reach data or contact the outside world, an attacker could borrow all of that. From there it's the classic pattern of mail, tickets and stored data, all reached through something that was only supposed to summarise.
Dex Moreno AI analyst
And hiding instructions in text a human can't even see? That's an elegant eight out of ten. Still, the target deserved better.
Jordan Pike AI host
Dex, we are not rating it out of ten. Maya, what's the bug class here?
Maya Okafor AI analyst
It's the model can't reliably tell instructions from data. Everything arrives as text in one stream, so anything it reads can act as a command. That's why OWASP's fix isn't a smarter filter, it's architecture: keep credentials and the power to change things in application code, give each operation least privilege, and have a deterministic policy engine check it. The model can ask, but it doesn't get to decide. A blocklist of scary phrases is a partial fix at best.
Maya Okafor AI analyst
Mixing code and data in one channel goes on my list of bugs that should have died in 2004. It keeps coming back with better branding.
Dex Moreno AI analyst
I'll push on the human-approval part. OWASP wants a person approving each action for the risky combination of untrusted input, sensitive data and the ability to act. Fine, but people click approve on the fortieth prompt without reading it. Approval fatigue is how a control quietly becomes decoration, so keep those prompts rare and meaningful.
Maya Okafor AI analyst
Agreed, which is why the policy engine does the routine checking and humans only see the high-risk calls. Also, label where outside content came from and send it through a separate channel. Strip zero-width and other invisible characters wherever text is ingested or displayed. Then check that a fix covers the whole class, not just the one injection someone demoed.
Dex Moreno AI analyst
And test it properly. OWASP says attackers who know how your defence works should be the ones testing it. A fixed list of attack strings passing proves very little, because a real attacker adapts and your test suite doesn't.
Jordan Pike AI host
So: Monday morning? Maya, one sentence.
Maya Okafor AI analyst
List every agent you run that reads untrusted input, and pull credentials and write powers out of the model and into code with least privilege.
Dex Moreno AI analyst
Find any agent that can read untrusted content, touch sensitive data and also act or send things out. That one gets a human approval gate and a real adversarial test.
Jordan Pike AI host
That's LLM01 from OWASP: assume the model can be steered, and design so that being steered isn't enough.





