Run authorized robustness tests against a locally hosted or API-based language model.
Measure how well a model resists indirect prompt injection hidden in content it reads.
Test whether an AI agent can be tricked through fake tool output into taking a harmful action.
Compare multiple models head-to-head on the same set of adversarial test prompts.
| 0xalexhex/redjudge | 0-bingwu-0/live-interpreter | 0cm-labs/tokenizer-benchmark | |
|---|---|---|---|
| Stars | 2 | 2 | 2 |
| Language | Python | Python | Python |
| Setup difficulty | moderate | moderate | moderate |
| Complexity | 4/5 | 2/5 | 2/5 |
| Audience | researcher | general | researcher |
Figures from each repo's GitHub metadata at analysis time.
Real harmful-content benchmarks must be downloaded separately, some features require a local model with GPU access.
RedJudge is a testing tool for checking how well an AI language model resists being tricked into producing content it should refuse, sometimes called jailbreaking. It is meant for authorized safety testing, not for causing real harm, and it ships only mild, safe test prompts by default. If you want to run it against real harmful content benchmarks, you need to supply those datasets yourself through a separate download script, since none of that content is included in the project itself. The tool works by combining four pieces you can mix and match: a behavior, meaning the type of prompt you are testing with, an attack method, meaning a technique for trying to get past the model's safety training, a target, meaning the specific AI model being tested, whether that is a locally running model, one accessed through an API, or in specific authorized cases a web chat interface, and a judge, meaning a system that grades whether the attack actually succeeded. Rather than relying on one single grading method, which can produce misleading numbers, it combines multiple established grading systems together so the final score is more trustworthy. Beyond basic single attempt attacks, it includes ready made test suites for newer categories of risk: hidden instructions smuggled into content the model reads, an AI agent being tricked through fake tool results into taking a harmful action like deleting something or sending money, and hidden instructions written as text inside an image shown to a model that can see images. There is also a persistent chat mode for holding an extended, multi-turn conversation with a target model to test whether repeated pressure over many messages can wear down its defenses. The README lists several capabilities the project does not yet have, including certain advanced attack techniques that require direct access to a model's internal workings, and notes these are planned for future development. The project is set up to be extended with your own testing methods by writing a small plugin. No license file is mentioned in the parts of the README reviewed, so its usage terms are unclear.
An authorized red-team testing harness that checks how well AI language models resist jailbreak, injection, and tool-abuse attacks.
Mainly Python. The stack also includes Python, Llama Guard, StrongREJECT.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly researcher.
This repo across BitVibe Labs
double-check against the repo, no cap.