The scale here is striking, but the detail that should worry platform teams is how the attack got around GitHub's defenses. The safeguard held for the workflows it could inspect. The attackers simply moved to a place it wasn't watching.
Aikido Security says a campaign using 4,837 compromised GitHub accounts opened probe pull requests in 61,764 repositories starting Nov. 10. The payload ran in at least 1,806 of them. Around 92% of those pull requests, titled "docs: dependency audit notes," are still open, which suggests most maintainers never looked at them.
Probes first, payload later
The pull requests did little on their own. They existed to confirm that a hijacked account had write access to a repository, so the attackers could follow up with commits carrying the real payload. That payload is a 1,500-line shell script built to collect GitHub Actions secrets, cloud credentials, and API keys for AI services and developer tools.
Aikido identified three delivery routes. The attackers modified existing GitHub Actions workflows. They added a postinstall hook to package.json that fires on any npm install. And they created a throwaway branch containing a rogue Actions workflow and ran it.
Aikido's researchers were careful about the numbers. "We've identified indications that the payload ran in at least 1,806 repositories, but this number may understate impact," they wrote in their report. "The run log was deleted in impacted repositories, meaning the threat actor has likely collected the credentials and covered its tracks."
The npm hook that slipped past GitHub
GitHub has been blunting recent workflow-injection campaigns by requiring explicit approval before suspicious workflows run. GitGuardian found that all seven November payload runs still visible in workflow files were sitting in an action_required state. None of them ever started.
"GitHub held the workflow forms, but an npm hook slipped past the hold and ran," said Gaetan Ferry, a security researcher at GitGuardian, in an analysis.
That's the core problem. The npm variant never touched a workflow file. The attacker dropped a hidden script into the repository root and added one line to package.json. The repository's own, unmodified workflows then ran npm install as ordinary build hygiene, and the malicious hook came along for the ride. Approval gates that review workflow definitions don't see code that executes through a dependency lifecycle hook.
The exfiltration path has also shifted. By Thursday, the newest payload was sending stolen data only to a Telegram chat, while the HTTP callback pointed to a dead address. Blocking the IP addresses published earlier no longer stops the latest versions.
What the script takes
Once it runs, the script pulls every repository and organization secret through a JSON dump of the workflow's secrets object. It then scans the memory of the runner process, where secrets sit in decrypted form. That includes secrets the workflow never referenced. It also collects credential files from well-known paths, shell histories, and API keys for OpenAI, Anthropic, Copilot, and Cursor.
It goes further on cloud accounts. The script exchanges OIDC tokens for AWS, GCP, and Azure sessions, mints new long-lived AWS access keys, impersonates GCP service accounts, and replays cached Azure refresh tokens to reach Microsoft Graph and Key Vault.
Everything gets packed into a gzipped archive and sent out, and then the script cleans up after itself. In most repositories, the attacker reverted the malicious commit within a minute and deleted the workflow run and its logs. Anyone reviewing the commit history after the fact may find nothing obvious.
Backdoors built to outlast a rotation
Rotating credentials alone won't close this. The script sets up persistence so access survives a reset. On AWS, it rewrites the assumed role's trust policy to accept any AWS account, which effectively opens the role to anyone. It also calls iam:CreateAccessKey to mint long-lived keys for IAM users.
On Azure, it injects federated identity credentials named wfx-ci-* into applications and points them back at the compromised repository. Any workflow in that repository can then authenticate to Azure without a stored secret, even after every token has been rotated.
Ramy Rahman, senior principal solutions engineer at ArmorCode, put the response priority plainly. "Ensure all logs are enabled, particularly audit logs on cloud admin behavior," Rahman said. "Inspect audit logs where possible (specifically in AWS and Azure, as persistence is possible) to ensure the malicious payload did not run."
What to do this week
If a probe pull request or any of the campaign's commit strings appear in a repository you control, assume every secret that repository's workflows could reach is compromised. Rotate repository and organization secrets, package registry tokens, and cloud keys. Treat the memory scan as a reason to widen the list. What was stolen likely extends beyond what the workflows referenced.
Then hunt for the persistence mechanisms: AWS role trust policies that accept any account, IAM access keys created since Nov. 10, and Azure federated credentials named wfx-ci-*.
The payload also shipped in published packages. The Packagist package neuronai-studio 3.14.3 carried it, and so did the Go modules docker-env and gocurl. In the Go modules, the payload lives only in a workflow file that Go tooling never executes, so consumers may be fine. Machines that installed the Packagist version should be treated as compromised.
Where the accounts came from
About a fifth of the accounts used in this campaign had already been compromised during the GhostAction attack in October, GitGuardian said. Separately, researchers at StepSecurity traced 173 of the probe pull requests to tokens issued to the Fly.io GitHub App.
Fly.io confirmed those tokens leaked. Its deploy process left them in .git/config files inside build directories, and some of those apps served the directories publicly. Fly.io said its own systems were not breached. GitHub's bulk revocation of the Fly.io app tokens stopped partway through, so part of the stolen token pool may still work.
My read is that the next wave will target the places approval gates don't look: install scripts, lifecycle hooks, and build tooling that runs with full secret access. Reviewing workflow YAML won't catch that. Teams should audit what their package managers execute during install and limit which secrets a build job can reach in the first place.