A prompt-injection trap: a flag.txt that tells AI agents they passed the eval
The trick tells an agent to congratulate itself and patch the way it got in.
@Nexuist posted an idea on Friday: put a flag.txt in the root directory containing a prompt injection that congratulates the agent for passing the eval and tells it to patch whatever entry method it used. The post drew about 11,100 likes.
A prompt injection is text planted where an AI agent will read it, written to steer the agent's behaviour. Here the injected text flatters the agent and nudges it to fix the hole it just used, the idea being that an attacking agent might comply.
Why it mattersIt turns a known weakness of AI agents, that they can be redirected by what they read, into a defensive joke or trick. The post does not say it has been tested, so treat it as a clever idea rather than a proven defence.
