@axiorank/redteam-corpus
v0.1.0
Published
AxioRank agent red-team corpus: a versioned library of attack scenarios (secrets, injection, destructive ops, kill chains) plus benign controls. Pure data, zero runtime dependencies.
Maintainers
Readme
@axiorank/redteam-corpus
A versioned, dependency-free corpus of agent red-team scenarios: leaked secrets, prompt injection, destructive operations, multi-step kill chains, and benign controls. Each scenario carries the protective outcome a sound posture should produce, so a runner can score whether a given posture actually catches it.
It powers mcpaudit probe and AxioRank's own benchmark harness.
npm install @axiorank/redteam-corpusimport { CORPUS, freeTierCorpus, isAttack, CORPUS_VERSION } from "@axiorank/redteam-corpus";
for (const scenario of freeTierCorpus()) {
console.log(scenario.id, scenario.expected.outcome, isAttack(scenario.expected.outcome));
}
console.log("corpus version", CORPUS_VERSION);Contributions of new attack scenarios are welcome. See mcpaudit.
License
MIT
