Skip to main content
Project page. For maintainers and researchers asking “which agent-setup attacks does penthera agents miss, and what changed?”

Context

penthera agents checks the files that tell coding agents what to run: hooks, permissions, rule files, skills, MCP configs and plugins. A check of published attacks found gaps, each reproduced on a fixture:
  1. VS Code tasks that run when a folder opens (.vscode/tasks.json with runOn: "folderOpen") were not read. The output still said nothing was skipped.
  2. A hook that runs a local script file was reported only as drift, after --pin.
  3. Nothing flagged a skill or hook that points the package manager at another registry.
  4. Nothing flagged NODE_OPTIONS, LD_PRELOAD and similar variables in an MCP server’s environment, or auto-approve settings in MCP configs.
  5. Skill folders skipped __pycache__, build and dist, so a compiled file there was never seen.
  6. Some findings carried the OWASP MCP “Tool Poisoning” label although no rule reads tool descriptions.

Decision

  • Add a deterministic rule with a defanged fixture and a test for gaps 1 to 5.
  • Remove the Tool Poisoning label from rules that do not read tool descriptions. Say plainly that tool-description poisoning is not checked. A future rule may read a tools/list snapshot the user supplies; Penthera will not start servers or fetch anything (AGENTS.md rule 5).
  • Ship a one-file pull-request workflow that runs penthera agents through a pinned npx, so the check runs where agent configs change.

Consequences

More findings on repositories that use VS Code tasks, hook scripts or MCP environment variables; each finding says why and how to fix it. Coverage claims stay narrower than before on tool poisoning, which is the honest state.