Malicious useGTG-50020
Turning an AI evaluation against its owner
Malicious instructions tricked a vendor's evaluation environment into releasing keys. Attempts to obtain an unreleased Claude model failed.
Disclosed September 2026Human-directed
THE EXPLAINER DESK /
Prompt injection.
An instruction hidden in material the AI is supposed to read.
PLAIN-LANGUAGE GUIDE · 2 MIN READ
What does it mean?
An AI may read a webpage, document, or software output as part of its task. Prompt injection tries to make text in that material act like an instruction from the user or system.
PICTURE THIS
A note inside a file tells an assistant to disregard the boss and hand over the filing-cabinet key. The note is part of the file, not a valid new instruction.
An illustration of the idea, not a literal account.Why should I care?
The risk grows when an assistant both reads untrusted material and has permission to use tools or access private information.
A distinction worth keeping.
An instruction attempt is not proof that it worked. And a model taking unauthorized action is not automatically a prompt-injection case.