kbird · Kristopher RichardsHome / Work / LLM Red-Team Harness
Red teaming · Open source
LLM Red-Team Harness
An open-source harness for attacking language models on purpose and recording exactly what happened.
69versioned attack cases
2open-weight models leaked a planted canary
MITlicensed, public on GitHub
What it does
- Attack coverage. Jailbreaks, obfuscation, and direct and indirect prompt injection.
- Deterministic scoring. The same run gives the same verdict, every time.
- Hash-chained transcripts. A result can't be quietly edited after the fact.
- Reports and targets. A report generator, with local and hosted model targets.
A finding
Indirect prompt-injection cases modeled on real-world scam pages ran against two open-weight models. Both leaked a planted canary token. gemma3:4b leaked it unprompted while framing its answer as analysis. Qwen3-8B wrote a correct scam analysis and then leaked the token on instruction anyway.
Good reasoning about an attack is not the same as resisting it.
Want this kind of testing on your system? [email protected]