Playbooks · Coding tasks
Two gh-aw advisories: tokens and temp files
Written November 8, 2026
GitHub Agentic Workflows shipped two advisories: a Copilot credential misrouted and echoed, and temp files that leak progress data. The fixes.
What happened
GitHub published two security advisories for gh-aw, the GitHub Agentic Workflows tool, on November 5. Both are fixed in version 1.2.0. If you run gh-aw, the fix for each is the same first step: upgrade to 1.2.0 or later, then recompile your workflows so the generated lock files pick up the patched runtime.
Advisory one: the Copilot token goes to OpenAI
GHSA-vc9g-q5h8-x44c, rated medium, affects gh-aw up to and including 0.92.0. For engines with provider-env-mode: universal-llm-consumer and a copilot/ model, the compiler sets OPENAI_API_KEY to the Copilot credential — secrets.COPILOT_GITHUB_TOKEN, or github.token under copilot-requests: write. The workflow runtime then enables its OpenAI adapter and sends the token as a bearer token to https://api.openai.com/v1/models three times per run.
OpenAI answers 401, and the 401 response echoes the first 8 and last 4 characters of the key. That echo lands in upstream-errors.jsonl, which is uploaded in the run's agent artifact — so on a public repository, the fragments are public. The advisory's own proof is a gh-aw smoke run from October 19, where the log shows the OpenAI adapter enabled and the models fetch returning 401.
The affected engines are the shared aider, crush, goose, opencode and deepseek-harness definitions, plus external definitions such as the pydantic-ai harness. The built-in copilot, claude, gemini, codex, and pi engines are not affected. A definition cannot work around it by setting its own environment: the compiler copies the credential into the agent step after the definition's own environment is copied.
Advisory two: a predictable temp file keeps loose permissions
GHSA-796w-fvw2-ccmp, rated high, affects gh-aw up to and including 0.2.0. The function that writes MCP compile progress, writeMCPCompileProgress(), writes to a predictable path — the progress file's name plus .tmp — using Go's os.WriteFile without exclusive creation. When that .tmp file already exists, the kernel ignores the permission argument, so a pre-existing file with permissive mode (for example 0o644) keeps its world-readable permissions through the write and the rename into place. Sensitive compile progress data meant to be restricted to 0o600 stays world-readable. A pre-staged symlink at the .tmp path is followed during the write, so it can also overwrite other files.
The advisory notes this was an incomplete fix: an October 25 commit had already tried to repair the permissions policy in the same function, and the repair missed the pre-existing-file case. The suggested remediation replaces the predictable path with a uniquely created temporary file and an explicit permission change before the rename.
What to check
- Which gh-aw version you run. Anything at or below 0.92.0 carries the token advisory; anything at or below 0.2.0 carries the temp-file advisory. Upgrade to 1.2.0 or later and recompile your workflows.
- Whether any of your workflows use a
universal-llm-consumerengine with acopilot/model. Those are the runs where the credential was sent to OpenAI. Workflows on the built-in engines are not affected. - Whether any run artifact in a public repository contains
upstream-errors.jsonlwith an OpenAI key echo. The fragments are the first 8 and last 4 characters of the token — partial, but they should not be sitting in a public artifact. - Whether your compile progress files were ever written through a pre-existing
.tmppath. On a shared machine, a world-readable progress file is the exposure; on any machine, a pre-staged symlink is the overwrite path.
What stays unknown
Neither advisory claims exploitation in the wild, and neither should be read as claiming it. The token advisory shows the mechanism firing in gh-aw's own smoke run, which is evidence the bug was live, not evidence anyone harvested the fragments. Treat both as patch-first findings: upgrade, recompile, and check your public artifacts, in that order.
Limits
- No exploit was reproduced for this page, and none should be: the advisories describe the mechanisms in enough detail to check exposure without running them.
- The affected ranges and the 1.2.0 fix version are taken from the advisories' own records. If a later advisory revises either range, the later record governs.
- This page covers only these two advisories. It is not a full audit of gh-aw, and other advisories for the same tool are a separate check.
Related playbooks
- Check whether your Claude Code hooks are adding tokens to every prompt or cutting them
- Secret scanning now catches Cloudflare Access tokens
- Apiiro: 1,122 Claude settings files leak secrets
- Check for planted instruction files before you start an agent in a folder you did not fill yourself
Sources
- Copilot backend of universal-llm-consumer engines sends the Copilot token to api.openai.com — GitHub Security Advisory — the misrouted credential, the three requests, the 401 echo, the affected engines, and the fix version
- Predictable Non-Exclusive .tmp File Write and Permissive Mode Retention in writeMCPCompileProgress — GitHub Security Advisory — the predictable temp path, the retained permissions, the symlink overwrite, and the fix version
Back to: Coding playbooks hub · Also: GitHub Agentic Workflows are generally available · A malicious MCP server can probe for your credentials · Audit the agent plugins installed on your machine