Vulnerability  ·  2026-11-21

pmo-platform Claude Code guardrail hooks silently skip egress/credential/MCP controls when jq is unavailable

VulnerabilityMedium impactGlobalCVE-2026-77350
The block-* hooks hard-code jq and return success when jq is unavailable at the pinned path, so the credential-read/egress/filesystem/MCP-write controls effectively no-op on affected macOS via Homebrew-only jq. Native deny rules partially backstop credential paths but not egress/MCP. CWE-693, CVSS 3.1 7.3.
A Claude Code agent's guardrails silently fail on a common toolchain setup, so a prompt or repo that would normally be stopped (credential read, arbitrary egress, MCP write) passes through — a guardrail-bypass in an AI coding agent's security boundary.
The guardrail hooks pin PATH to /usr/bin:/bin and hard-code jq as /usr/bin/jq, returning success when jq is absent from that path, so an attacker who can influence a prompt/repository/evaluation flow can cause the model to read credentials, make network egress, write outside the filesystem boundary or perform MCP writes without interception.
pmo-platform < 3.69.1 (GHSA-9cjm-v22x-4x33)
Upgrade pmo-platform to 3.69.1 (resolves jq lookup robustly) or ensure jq is present at /usr/bin/jq on the affected system.
GHSA-9cjm-v22x-4x33Fix commit (586b968)NVD record
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →