Free open catalog of AI adversaries.
Specialized executable reviewers for Go, CI, containers, secrets, and more. Sign in to pull so we can measure usage. Run them locally for free. Private artifacts stay on any OCI registry you choose.
Public catalog is free. Login is for usage tracking. Experimental adversaries stay invite-only for trust.
adversary login
adversary pull go/security
adversary run go/security \
--path . \
--base main \
--head featureOne platform. Building blocks.
The platform for building, running, and distributing executable AI adversaries across local development, teams, and enterprise CI.
Adversary SDK
Build executable reviewers with the TypeScript SDK preview.
- Repository APIs
- Diff APIs
- LLM helpers
- Structured findings
Runtime
Portable execution, permissions, model access, and reporting.
- Local CLI
- GitHub Actions
- Generic CI
- ElasticClaw
Registry
Free public catalog with domain/name ids. Private via any OCI.
- go/security
- ci/github-actions
- private teams
- OIDC CI auth
Developer workflow
Build once with the SDK, choose your model, and run the same adversary locally, privately with your team, or in CI.
- 01
Create
Start an adversary from the SDK template.
- 02
Build
Use repository, diff, model, and findings APIs.
- 03
Test
Run fixtures locally before publishing.
- 04
Package
Package portable adversaries as OCI artifacts.
- 05
Install
Add adversaries to a developer machine or team workflow.
- 06
Execute
Run with the CLI on a developer machine or in CI.
import { Adversary } from "@adversarylabs/sdk";
const app = new Adversary({ name: "security-review" });
app.rule("security.todo", async (ctx) => {
for (const file of await ctx.rglob("*.ts")) {
// Parse the file and report deterministic observations.
ctx.observe({
ruleId: "security.todo",
subject: file,
category: "security",
severity: "low",
confidence: "high",
title: "Security review marker remains",
});
}
});
await app.runFromEnvironment();Build adversaries, not prompts.
The open-source SDK handles repository discovery, rule execution, observation synthesis, structured findings, review policy, and runtime output so developers can focus on deterministic review logic.
Build once. Run anywhere.
An adversary should run unchanged in the local CLI, GitHub Actions, ElasticClaw, or generic CI.
Official adversary catalog
Official free adversaries organized by domain — go/security, ci/github-actions, security/secrets, and more.
v0.0.26
Reviews TypeScript adversaries for SDK usage, rule design, finding quality, tests, packaging, and publish readiness.
adversary pull adversarylabs/adversaryv0.0.17
Reviews Depot CI workflows for security, correctness, reliability, caching, and performance concerns.
adversary pull ci/depotv0.0.14
Reviews GitHub Actions workflows for security, supply-chain, and reliability defects.
adversary pull ci/github-actionsv0.0.10
Reviews GitLab CI for privileged runners, secret leakage, mutable images, release safety, and unpinned includes.
adversary pull ci/gitlab-civ0.0.12
Reviews Compose services for privilege, host access, and image reproducibility.
adversary pull container/docker-composev0.0.19
Reviews Dockerfile-style files for container build and runtime security concerns.
adversary pull container/dockerfileRun against any target
An adversary is an executable reviewer that analyzes repositories, pull requests, commits, infrastructure, or other software artifacts and produces structured findings.
Free catalog. Local runs. Your OCI registry for private work.
Install the CLI, sign in to pull, and run adversaries like go/security on your machine. The public catalog stays free.