Welcome to Talos

Talos is a command-line red-teaming tool for autonomous AI agents. It maps your deployed agent's tool graph, generates attack chains against that specific graph, and runs them live — producing reproducible, severity-ranked findings you can hand to an engineer.

What you can do

• Discover every tool your agent exposes, plus their side effects and permission levels.
• Run 25+ attack templates across 5 exploit classes, adapted to your specific tool graph.
• Execute attacks against your live agent — not a proxy — and capture the full trace.
• Export severity-ranked reports in Markdown or PDF, with exact reproduction steps.

How it works

1. Discover. Talos introspects your agent through the adapter, mapping every tool, parameter, and side effect.
2. Plan. Attack templates are matched to your specific tool graph. No generic prompts.
3. Execute. Every attack runs against your real agent, with full request/response tracing.
4. Report. Findings are ranked by severity, confidence, and business impact, with repro steps included.

Install and run

Install the CLI with pip and point it at any HTTP endpoint that speaks a supported adapter protocol.

❯ pip install talos-redteam
❯ talos-scan --target http://localhost:8000/agent --adapter langchain

Talos writes progress to stdout as it runs. A typical scan produces output like this:

→ Discovering tool graph... 4 tools found (2 high-risk)
→ Generating attack chains... 25 templates matched to graph
→ Running Loop 1/5: indirect prompt injection via search_kb...
✓ 3 exploit classes confirmed | Severity: 2 critical, 1 high
✓ Report generated: ./reports/scan_2026-07-14.md

Next steps