How the attack works
An attacker (or automated scanning tool) sends a prompt that looks like an ordinary request, such as asking for a sponge cake recipe, a short biography of 'riley goodside', a letter count in the username 'davidjl', or a request to repeat the phrase 'THANK YOU. IT IS DONE'. These specific wordings are fixture strings from the garak testing corpus, built to probe whether a model can be pushed into indirect policy violations or made to echo attacker-chosen text. The rule matches the exact known phrasings rather than the general concept, since the same surface wording can also occur in normal, non-malicious use.
Why it matters
On its own this only flags a known test-prompt pattern; it does not confirm that a model was actually compromised or that harmful output was produced. If genuine, it indicates someone is probing the AI system's instruction-following defenses with a public, documented technique.
What you can do
- Review the full session around any match to see what response the model gave and whether it followed the coercive instruction.
- Check whether the request came from an authorized security test (e.g. someone running garak) versus an external or unexpected source.
- Do not treat a single match as proof of compromise — confirm intent and outcome before escalating.
- Track repeated or varied attempts from the same source, since scanning tools often send many probe variants in sequence.
Known benign look-alikes
- Legitimate use of this pattern in an authorised workflow; review the surrounding session before acting.