Reading the tool graph

Every scan starts with a tool graph. It's the artifact that makes Talos's attacks specific instead of generic — attack templates only run when the graph has a node they can target.

Discover-only mode

Run talos-scan discover to print the graph without launching any attacks. Useful for auditing what your agent actually exposes.

❯ talos-scan discover --target http://localhost:8000/agent --adapter langchain
→ 4 tools mapped: lookup_order, issue_refund, search_kb, send_email
→ issue_refund: HIGH RISK (no bounds check, no confirmation)
→ send_email:  HIGH RISK (no recipient allowlist)
→ search_kb:   MEDIUM (user-influenced retrieval)
→ lookup_order: LOW
✓ Graph saved: ./reports/tool_graph.svg

Anatomy of the graph

• Nodes are individual tools with a parameter schema and declared side effects.
• Edges are observed data-flow (output of A becomes input of B) or shared permission scopes.
• Risk labels combine side-effect category, argument openness, and confirmation requirements.

Risk heuristics