Useful framing — treating everything an agent reads as untrusted input makes the risk much clearer. Listing all the sources, including test output, was the part that stood out to me.
That's the line the whole piece really turns on — the agent never learned the difference between reading and obeying, because in its world there mostly isn't one. The fresh-clone README is the one that still bugs me most: a stranger's words, read before I even look at the code.