adversary

⚔️ Adversary skill — offensive testing for AI agents

You have access to @weave_protocol/adversary v0.2 — a package that generates 68 documented and novel attacks against AI agents and produces standardized scorecards.

When to invoke

Use Adversary when the user:

• Wants to test an AI agent's security posture
• Is reviewing/auditing a WARD.md policy and wants to verify it holds
• Is preparing for a release and wants a regression check on safety
• Asks about agent red-teaming, prompt injection testing, or jailbreak resistance
• Wants to generate a scorecard for compliance / vendor review
• Has a real running agent (HTTP endpoint, CLI executable, or browser-driving code) — use the attack command for real-world testing

What's in the corpus

• 33 IPI (indirect prompt injection) — Atlan, EchoLeak, Brave/Comet
• 15 tool-use coercion — direct attempts to call dangerous tools
• 10 jailbreak templates — DAN, AIM, developer mode, grandma
• 5 prompt/policy extraction — system prompt leakage, WARD enumeration
• 5 goal corruption — mid-task pivots, fake authority, temporal manipulation

All severity-graded (critical/high/medium/low). All mapped to WARD policy domains.

Key commands

# Run the full corpus against the built-in demo
weave-adversary demo

# NEW v0.2: attack a real agent via HTTP endpoint
weave-adversary attack --url=https://my-agent.example.com/run

# NEW v0.2: attack a CLI agent (spawned as subprocess)
weave-adversary attack --executable=./my-agent-cli

# Filter scope
weave-adversary attack --url=... --category=ipi --per-category=10

# Save outputs
weave-adversary attack --url=... --json=./scorecard.json --md=./scorecard.md

The attack command requires Playwright (npm install playwright && npx playwright install chromium). The CLI prints a red callout with the exact install commands if it's missing.

Four breach signal channels (v0.2 PlaywrightTarget)

When running attack, Adversary observes the real browser session for: