Hijacked GitHub accounts opened probe pull requests in 61,764 repositories for the GitHub secret stealer known as wfx in November 2026, and the campaign was still opening new probe pull requests on the morning of 12 November. Aikido Security, which first reported the campaign on 12 November, counted 4,837 compromised accounts opening probe pull requests in 61,764 repositories, and found the payload ran in at least 1,835 of them. BleepingComputer confirmed the same figures on 12 November.
The attack, which Aikido calls wfx-loot, is not a flaw in GitHub Actions. StepSecurity, which calls the malware WFX Phantom, says the attacker writes code with legitimately issued but stolen tokens, and no vulnerability in GitHub Actions itself is being exploited. What makes it dangerous is what the payload does once it runs: it collects the workflow’s secrets, tokens and credential files, plants persistence built to survive rotation, and wipes the logs that would prove it was there.
How do I check if the GitHub secret stealer hit my repos?
Search every repository you can write to for a branch named wfx-docs, a pull request titled docs: dependency audit notes, and any gap in your workflow run numbers since 10 November. A probe pull request is evidence the account that opened it was abused, so investigate it at once and move to rotation.
- Search every repository you can write to for a branch named wfx-docs from the 10 November campaign.
- Look for any pull request titled docs: dependency audit notes opened from a wfx-docs branch.
- Scan your workflow run history for missing run numbers, the usual sign that logs were deleted.
- Check for short-lived chore/ branches with eight random characters and ci-check- branches with four digits.
- Check commit history for ci: refresh build cache, chore: add ci helper, chore: add dependency audit workflow, ci: restore and ci: cleanup.
- Compare the stealer script hash 89695abb392d5cfd4972672961e644bb46baa98aa1ed7173c2a2816c762735cf against every workflow file in your repositories and investigate any match at once.
A probe pull request opened from your account is evidence that the account’s credentials were abused, so rotate them even if no payload followed. The probe tests whether the account can write to the repository, and the attacker can introduce the payload where the account has that access.
What does the GitHub secret stealer take?
The payload is a single shell script of about 1,500 lines, which labels itself wfx in its logs and variable names, according to Aikido. It collects GitHub Actions secrets and variables, the workflow GITHUB_TOKEN, OIDC and Actions runtime tokens, AI-service keys, and secrets from runner process memory. It also reads credential files for AWS, Azure, GCP, Docker, Kubernetes, npm, PyPI, Git, SSH, GPG, Vault and Terraform.
The stolen data first went by HTTP POST to 194.150.166.244:8080/exfil or 18.143.168.96:8080/exfil. A third server, 153.75.82.159:8080, appeared later, according to SafeDep as reported by The Hacker News, and by early on 12 November the operator had switched new payloads to Telegram-only exfiltration, according to StepSecurity and Corgea. About a fifth of a 200-payload-commit sample sent data only via Telegram, according to SafeDep.
The payload is also built to survive rotation. It can rewrite an AWS role’s trust policy to “Principal”: {“AWS”: “*”}, widening it to any AWS principal, and it can add federated identity credentials named wfx-ci- to Azure AD applications, which are built to survive rotation.
How the attack unfolds in four moves
Every attack starts with a probe pull request titled docs: dependency audit notes, opened from a branch named wfx-docs. The probe tests whether the compromised account can write to the repository. Where it can, the attacker pushes the stealer into a workflow, the payload runs and sends credentials out, and then the attacker reverts the change, usually within about a minute, and deletes the workflow run and its logs. A missing run number in the workflow’s history is often the only trace left.
The campaign started at about 12:45 UTC on Tuesday 10 November. By 15:00 UTC on 11 November, GitHub search returned 61,839 probe pull requests. SafeDep counted 8,093 repositories that received malicious commits between 12:46 UTC on 10 November and 04:36 UTC on 12 November, and the campaign was still active when it searched. SafeDep found the stealer added to an existing workflow in 7,309 repositories and delivered as an npm postinstall hook in 771. The Hacker News found fresh probe pull requests early on 12 November, three of them between 05:04 and 05:38 UTC from wfx-docs branches. StepSecurity counted more than 63,000 probe pull requests by 07:40 UTC on 12 November, about 90 percent still open, with new ones appearing every few minutes.
Only about 3 percent of the probed repositories actually ran the stealer, our own calculation from Aikido’s figures: 1,835 runs against 61,764 probed repositories. Aikido’s post headline gives 1,835; its body still carries the earlier count of 1,806. Against SafeDep’s count of repositories that received a malicious commit, the share is higher: about 23 percent of the 8,093 ran it. The two counts come from different researchers, so this is a comparison of their figures, not one measured rate. The gap between probed and ran is the story, by our reading of the counts. The probe was the mass event; the run was the selective one.
What should I rotate if the payload ran?
Rotate every secret and cloud role available to the affected workflows, then hunt the persistence the payload plants: an AWS role trust policy opened to everyone and federated credentials named wfx-ci- in Azure. Rotation alone is not enough, because the payload is built to survive it.
- Rotate every secret and variable available to the affected workflows, not just the ones you use.
- Revoke and reissue the workflow GITHUB_TOKEN, OIDC tokens and Actions runtime tokens for every affected repository.
- Audit every AWS role the workflows could reach for the wildcard trust-policy change and remove any malicious policy found.
- List every federated identity credential in your Azure applications and investigate any named wfx-ci-, removing confirmed malicious additions at once.
- Check every machine that installed or opened Packagist digitalelvis/neuronai-studio version 3.14.3 and treat each one as compromised as a precaution.
- Remove Go module github.com/onuragtas/docker-env versions v1.0.278 and v1.0.280 from every build and deployment pipeline now.
GitHub published advisory GHSA-j3hx-5mxx-43qh at 11:33 UTC on 12 November, and npm unpublished pdf-html5-page-flip version 2.0.2, according to Corgea. That package is tied to the overlapping October GhostAction wave, according to Corgea, not to wfx itself, but the two campaigns have overlapping compromised accounts, so check it in the same pass.
Is wfx the same attack as GhostAction?
No, not according to Aikido. The compromised accounts overlap with those hit in the October GhostAction wave, in which stolen GitHub accounts pushed secret-stealing workflows into repositories. Accounts hit in October were about 1,000 times more likely than other active accounts to be used in November, and in more than 100 repositories the November payload was injected into the October workflow files, leaving the October exfiltration step in place. Aikido does not attribute the November campaign to the same actor, saying infrastructure, tooling and tradecraft differ.
What is still unknown
Aikido does not explain how the attacker obtained the stolen GitHub credentials, and none of the reports read explains how 4,837 accounts were taken over. Aikido’s figures cover only the public repositories the company could see, so the true totals are higher, and the payload’s settings suggest it may also be running in private repositories, though that is not confirmed. GitHub had not commented on the campaign in any report read as of 13 November.
At least 50 of the compromised accounts publish packages for npm, Go, Packagist, PyPI or crates. Aikido confirmed the payload reached at least one Packagist package and one Go module through automated publishing triggers. Packagist package digitalelvis/neuronai-studio version 3.14.3 is malicious, and Aikido advises treating any machine that installed or opened it as compromised; version 3.14.4 fixes it, so update to 3.14.4. Go module github.com/onuragtas/docker-env versions v1.0.278 and v1.0.280 carry the payload in a GitHub Actions workflow file only; the Go source and release binaries are unchanged, and Go builds never run workflow files, but Aikido still recommends removing these versions. SafeDep also found the stealer in Go module github.com/vkuznet/gocurl, in a workflow that has never run, and in 28 WebdriverIO repositories through commits authored as the bot account wdio-bot, where no package release was affected but the bot’s token should be treated as stolen.
And the exposure window was still open when Aikido checked. About 92 percent of the 10,431 probe pull requests that went into repositories the compromised account does not own were still open when Aikido checked on 12 November, which is about 9,597 open malicious probe pull requests by our own calculation. If you own or write to any repository, start with the check.


