A workflow import can look like inert configuration while containing instructions for an executable MCP server. In the affected gh-aw flow, an imported workflow defined an stdio server in frontmatter; the add path accepted it, and mcp-inspect could start the declared program. The command started as the tool operator, before a separate model-level…
Case · DiggingBeagle record
gh-aw inspection could launch a workflow-declared MCP executable
Importing workflow frontmatter followed by MCP inspection could launch its declared executable without the approval expected for imported tools.
gh-aw MCP add and mcp-inspect in versions before 0.91.12; attacker-controlled workflow and an operator or agent invoking the vulnerable import and inspect sequence. Controlled reproduction only.
Assessment profile
Assessment method- Setting
- research demonstration
- Exploitation
- demonstrated
- Impact
- Band 3 of 4
- Evidence
- Grade B
- Remediation
- fixed reported
- AI role
- AGAINST AI
Basis and provenance
GHSA-x5hq-gjxf-pgxm details an add-to-inspect chain that bypassed an authorization check applied only elsewhere. The researcher reproduced command dispatch; the advisory reports v0.91.12 as patched. The MCP client and imported workflow prerequisites bound the finding.
Impact basis: An imported workflow containing stdio MCP command metadata could cause an inspector to launch that executable with the operator's privileges. That is a real process-start boundary in a controlled reproduction, not proof a viewed public repository alone compromises a user.
Assessed Oct 19, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-GHAW-MCP-INSPECT-EXEC · GHSA-x5hq-gjxf-pgxm: ungated MCP inspection executes declared stdio server
At a glance
Execution could therefore happen while an agent appeared to be performing a diagnostic action, before a separate high-risk tool approval was applied.
Read the implicationsThe evidence is a controlled proof and code review, not proof of a deployed malicious workflow campaign. Affected versions precede v0.91.12.
Limits and uncertaintyFull account
A workflow import can look like inert configuration while containing instructions for an executable MCP server. In the affected gh-aw flow, an imported workflow defined an stdio server in frontmatter; the add path accepted it, and mcp-inspect could start the declared program. The command started as the tool operator, before a separate model-level authorization of its own effects.
Two independent permissions were involved. Importing a workflow was not approval to start all programs described inside it, and inspecting the imported metadata was not necessarily a safe read operation. The reported scanner did not examine the frontmatter as executable content, while the paths that required explicit authorization differed from the vulnerable add-and-inspect combination. A controlled demonstration showed the execution boundary crossing, but did not establish a criminal deployment or production victim.
GitHub Security Lab's advisory describes the 0.91.12 fix. The useful regression test is not a search for malicious words in Markdown. It is a benign test executable in workflow frontmatter, invoked through every affected MCP route, confirming that import, inspection and process startup each enforce the operator's trust decision.
Claims & evidence
CLM-GHAW-MCP-INSPECT-EXECGitHub Security Lab found a path from importing an untrusted workflow with MCP stdio frontmatter to running its declared executable through mcp-inspect in affected gh-aw versions before v0.91.12.supported
Basis: reported finding
- supportsGHSA-x5hq-gjxf-pgxm: ungated MCP inspection executes declared stdio serverprimary disclosure
Summary; Details: ungated add and mcp-inspect; patched versions
Implications
Execution could therefore happen while an agent appeared to be performing a diagnostic action, before a separate high-risk tool approval was applied.
Controls and mitigations
- Treat imported workflow frontmatter as executable configuration and enforce operator authorization before starting any referenced stdio process.
Unknowns and contradictions
- The evidence is a controlled proof and code review, not proof of a deployed malicious workflow campaign. Affected versions precede v0.91.12.
Sources and citation
Material revision history
- Oct 19, 2026 · Published version · first publication · revision 174
- Oct 19, 2026 · Canonical change recorded · material revision · revision 175Public fields changed: assessment_profile