Skip to content

Security: loonylabs/dsh-pathfix

Security

SECURITY.md

Reporting a vulnerability

info@loonylabs.dev, or GitHub's private vulnerability reporting if enabled on this repository.

Please do not open a public issue for an exploitable security defect. There is no bounty and no SLA — reports will be reviewed and answered promptly.

What is worth reporting

dsh-pathfix intercepts tool execution parameters directly before tool dispatch. Because it modifies input arguments, safety defects in the normalization pipeline matter:

  • Regular Expression Denial of Service (ReDoS): Any input that causes unbounded backtracking or stalls the event loop in normalizePath or normalizePattern.
  • Path-traversal or boundary escape: Any normalization behavior that alters a relative or restricted path such that it circumvents a tool's intentional filesystem boundary or permission check.
  • Argument corruption: Any defect where normalization corrupts valid tool parameters or injects unexpected properties into tool execution contexts.

What is not a vulnerability here

  • Normal tool filesystem access: dsh-pathfix only repairs whitespace and trailing newlines on parameters passed by the model; whether the tool is allowed to read or write a given file is governed by the host harness filesystem security policy (ctx.fs, ctx.approval, or container sandboxes).
  • Model hallucination or invalid tool arguments: If the model provides a completely nonexistent file path, the tool will fail with standard filesystem errors; dsh-pathfix does not invent files or suppress genuine error returns.
  • Misconfigured profile layers: An upstream profile that excludes dsh-pathfix or disables it via enabled: false is operating as configured.

Please do not send

Live session logs containing API keys, credentials, or proprietary prompt text. Reproduce the defect using a minimal synthetic argument fixture in test/unit.test.mjs.

There aren't any published security advisories