Agent security testing thread: capture evidence before deleting the dependency graph
krishnan · x · 2026-11-18
Part 6/8 of @krishnan's thread on agent security testing. He proposes a test worth running: pinned dependency trees with known indirect vulnerabilities, including two versions of the same package, compared against an independently checked reference. Require a dependency path and an advisory for every finding, and measure setup time and missed findings.
This builds on part 5/8: deletion is the step to design. Before the graph goes, capture pinned inputs, queries, dependency paths and advisory evidence — the launch post describes a report but not what backs each finding, and deleting the database does not preserve that evidence.
More from coding & agent
- swyx after a week with Devin Bots: multi-agent teams that own outcomes, not tasks — marvinvonhagen · 2026-11-18
- AI Engineering from Scratch: first-principles guide to the agent loop and context budgeting — ghumare64 · 2026-11-18
- Vivek Haldar: hundreds of tiny enterprise tools are now cheaper to vibe-code than to maintain — vivekhaldar · 2026-11-18
- Adaption ships as a Hermes Agent plugin for in-agent custom model training — sarahookr · 2026-11-18
- Throwaway analysis environments: agents should leave reproducible evidence, no running servers — krishnan · 2026-11-18
- mathregistry Open Beta: An npm-Style Lean Math Registry with MCP Tools for AI Agents — DevinWatson · 2026-11-18