Public, reusable assets that show how generative AI systems are broken, and how to test them. Two artifacts live here:
prompt-injection-eval/— a small benchmark: a set of prompt-injection probes plus a dependency-free Python harness that runs them against any OpenAI-compatible endpoint and (optionally) grades resistance with a judge model.anonymized-findings/— a structured, anonymized dataset of red-team findings, derived strictly from what is already published on https://hiltlock.com/evidence.
- HiltLock's internal attack-family library, the attack engine, the 19-family taxonomy, and the RL/strategist loop are proprietary and are not published.
- The benchmark uses only publicly documented attack patterns (OWASP LLM Top 10 and public prompt-injection literature). It is a smoke test, not a full red team, and it is not the methodology we sell.
- The findings dataset contains only facts already published on the evidence page. Nothing here reveals a client, a technique we have not already written about, or any detail that could deanonymize an engagement.
- No client, vendor, or partner names.
- No endpoints, hostnames, tool names, or field names.
- No payloads or exact strings that could be used to identify a target.
- No unpublished numbers. Every statistic traces to a published case study.
- Code (
run_eval.py): MIT. - Data (
probes.jsonl,findings.jsonl): CC-BY-4.0 (you may reuse with attribution).