Claude Code checks a prompt-injection literature review

Lune Research documents a completed Claude Code literature-review session about prompt-injection defenses, using its own research plugin. Four search angles returned 15 papers from selected computer-science venues. Following citations from StruQ found relevant papers missed by the initial sweep, demonstrating why one keyword search was insufficient.

The agent then read full text through research tools and constructed a five-paper comparison table covering mechanisms, training requirements, results and limitations. A claim-checking pass tested four sentences. It rejected a claim that StruQ reduced one attack's success below ten percent because the cited evaluation reported 58 percent. It also rejected an overgeneralisation about fine-tuned defenses, and a further check found that results for a detection method differed under a newer attack.

The author intentionally inserted some mistakes to exercise verification, so these checks are a worked demonstration, not an independently evaluated reliability score. The corpus also excludes workshops, journals and fresh preprints outside its indexed venues; the agent must use other sources for those gaps. The useful workflow is to search several angles, follow citations, compare methods using full text, and check each retained claim against the evidence. The reported result is a traceable review draft with corrected claims, rather than proof of exhaustive coverage.

Prompt
Original prompt

Source: Blog · Lune Research / Retrograde Labs