Ask Astra to write a security audit prompt and it adds "do not invent CVEs"

morqon · x · 2026-10-26

A user asked Google's Astra to write a security audit prompt, and the model inserted its own instruction: "do not invent CVEs."

The line is telling — the model apparently knows it's prone to hallucinating plausible-looking but nonexistent CVE identifiers in security contexts, so it guards against itself. A knowing wink to anyone who has watched LLMs fabricate vulnerability IDs.

Original post →

More from Fun

Fun channel →