Your Pull Request Title Is Executable Input

Most security reviews of defense and space software start at the deployed system: the flight computer, the ground station, the radio link. The code gets there through a build pipeline, and that pipeline holds cloud keys, signing keys, and registry tokens. This week Radware published an analysis of an attack on those pipelines that needed no access to the target’s code.

A title field that ran as a command

Radware’s write-up, published October 29, describes activity on September 1. The researchers verified 83 malicious issues, pull requests, and comments across 61 repositories, 77 of them posted within about 45 minutes. The attacker did not need write access. Opening an issue or a pull request was enough, provided the project’s automation handled the text unsafely.

The weakness is a familiar one. Some GitHub Actions workflows insert values such as an issue title directly into a shell script, using the ${{ github.event... }} syntax. The shell then expands whatever the title contains. Radware’s example payload put a command substitution in the title that collected the job’s environment variables, encoded them, and posted them to an outside server. A payload sitting in a title does nothing by itself. It runs only when a workflow pastes it into shell source.

Radware analyzed two cases from run logs. In one, a pull request title on the PX4-Autopilot repository was passed to a title-checking script inside a Bash command, and the logs showed a null-byte warning and binary output reaching the script. In the other, an issue notification workflow built a variable from the issue title, and the generated script produced a Bash warning and a malformed message to Discord. Those are signs the injected text was interpreted. The authors are careful about what they did not establish. They did not confirm that any secrets were exposed, that source code was altered, or that any downstream software was compromised, and they describe the supply-chain impact as hypothetical.

GitHub’s own hardening guide describes the fix. Pass the untrusted value through an intermediate environment variable, so it never becomes part of the generated script, and quote it when the script uses it. It also advises making the default token permission read-only and raising it per job.

What a stolen pipeline credential reaches

The September campaign may have yielded nothing. The earlier Trivy incident shows what a stolen credential can do. Snyk’s account says attackers first abused a vulnerable pull_request_target workflow in the Trivy repository to steal an organization-scoped personal access token. In March they used access gained that way to rewrite 75 of the 76 version tags on the trivy-action repository so that they pointed at malicious code. A pipeline that referenced the action by tag ran a stealer before it ran the legitimate scan, so the build looked normal. According to Snyk, the stolen material on self-hosted runners included SSH keys, cloud credentials, kubeconfig files, and Terraform state.

Snyk’s main recommendation is to pin actions to a full commit SHA instead of a tag, because a tag can be moved. CISA’s May bulletin on the Megalodon campaign, which injected workflows into public repositories to harvest CI/CD secrets, adds that teams should pin software to specific trusted versions and audit workflow files for changes from automated-looking accounts.

Our reading: why this is a defense problem

This section is interpretation. None of the sources make these claims.

A flight software repository may use the same workflow patterns as any open project: a title check, a notification bot, a labeler, and now often an AI agent that reads issues and pull requests and posts comments. Each of those reads text written by someone outside the team and runs with the repository’s credentials. For an AI reviewer the situation is the same one as the shell script. The agent’s input is untrusted, and the question is what it can do with the token it holds.

Assurance work on defense programs usually treats the build as trusted infrastructure. These incidents suggest it is attack surface, and its credentials often reach further than the code it builds. Signing keys are the clearest example, since a team that can sign firmware can ship firmware.

What to check this week

Search your workflows for ${{ github.event inside any run: block, and move each hit into an env: entry. Find every pull_request_target and workflow_run trigger and confirm none of them checks out code from a fork. Replace tag references on third-party actions with full commit SHAs, and set the default token permission to read-only. Check whether any bot or AI agent that reads issues has write access or secrets it does not need, and whether signing keys are available to jobs triggered by outside contributors.