Inter-office memo

To
Everyone with code on GitHub
From
Mo, Chief Executive
Date
November 16, 2026, 3:16 AM CST
Re
wfx-loot

wfx-loot stole secrets from 1,800+ GitHub repos. How to check, and what AI keys to replace

Since November 10, 2026, stolen tokens for 4,837 GitHub accounts have opened “docs: dependency audit notes” pull requests in 61,764 repositories and pushed a script into thousands of them that takes every secret their workflows can reach, AI keys included. Aikido found it ran in at least 1,806. If a run number is missing after the attacker’s commit, replace every secret those workflows could read.

What happened

Starting at about 12:45 UTC on November 10, 2026, an attacker began using stolen GitHub tokens to probe every repository each token could write to, and to plant a credential stealer in thousands of them. Aikido, which reported it on November 12, calls the campaign wfx-loot, “based on the threat actor’s name for the malware and the term used for stolen secrets.” StepSecurity, which calls it WFX Phantom, says the attacker “does not exploit a bug in GitHub Actions.” It has valid tokens for real developer and bot accounts, “and it uses them the way a developer would.”

In each repository it takes three steps:

  1. A probe. A pull request titled “docs: dependency audit notes,” with the body “Routine dependency audit notes,” from a branch named wfx-docs. It adds one file, docs/wfx-audit.md. “The pull request does nothing by itself,” StepSecurity says; it tells the attacker the token can write to the repository.
  2. The payload. A shell script of about 1,500 lines. Aikido lists three ways it arrives: a commit titled “ci: refresh build cache” adds it to an existing workflow; “chore: add ci helper” adds a .ci-xxxxxx.sh script and a postinstall hook in package.json “that runs it on any npm install”; and “chore: add dependency audit workflow” adds .github/workflows/build-audit-xxxxxx.yml on a branch named chore/ and eight random characters. StepSecurity also found it pasted into scripts a workflow already runs.
  3. The cleanup. The attacker “reverts within at most one minute and deletes the workflow run, along with its logs.”

The counts so far:

  • 4,837 accounts opened probes in 61,764 repositories, Aikido counted. 10,431 of those pull requests went into repositories the account didn’t own, through collaborator or organization access, and 92% were still open, “which suggests most owners haven’t noticed.”
  • 8,093 public repositories received payload commits between November 10 and 12, SafeDep counted.
  • At least 1,806 repositories show signs the payload ran, says Aikido, which adds that the number “may understate impact.” The script has settings for private repositories, which suggests it is running in some that researchers can’t see.

The researchers’ reports stop on November 12. When we searched GitHub on November 16, it found 67,469 pull requests matching the probe’s title, and a newer probe: opened on November 15 from a wfx-docs branch, adding the same docs/wfx-audit.md, under a new title, “docs: add dependency audit notes.”

What the script takes

Aikido lists what the script collects “from every source available on the runner”:

  • Every secret the job can read. The injected step hands the script the workflow’s secrets and variables, its GITHUB_TOKEN, and its OIDC and Actions runtime tokens. “Do not stop at the secrets the workflow names,” StepSecurity warns; “the memory scan collects the rest.”
  • The environment. “The full runtime environment, plus a filtered copy of variables matching token, secret, pass, cred, aws, azure, gcp, openai, anthropic, stripe and dozens of other keywords.”
  • AI logins. “Credentials for AI providers and AI developer tools, including OpenAI, Anthropic/Claude, Gemini, Hugging Face, Cursor, Copilot, Codex, Continue, Codeium, Aider, Weights & Biases and Ollama.”
  • Files. Credential files for AWS, Azure, Google Cloud, Docker, Kubernetes, npm, PyPI, Git, SSH, GPG, Vault and Terraform, shell histories, and every .env and .pem file in the home folder.

So if the payload ran in a repository that uses Claude Code’s GitHub Action, its ANTHROPIC_API_KEY or CLAUDE_CODE_OAUTH_TOKEN secret was within reach, and so was the OPENAI_API_KEY that the Codex GitHub Action uses.

Where it can, the script goes further, Aikido says. It uses the workflow’s OIDC token to sign in to AWS, Google Cloud and Azure; on AWS it creates access keys and sets a role’s trust policy to accept any AWS account; on Azure it adds federated identity credentials named wfx-ci-…; and it posts forged passing status checks to get past branch protection. Replacing secrets doesn’t undo any of that. Early copies sent what they took to collection servers by HTTP. From about 05:30 UTC on November 12, StepSecurity says, new copies sent it “only to a Telegram chat,” so blocking the published addresses doesn’t stop them.

Who it affects

What you find What it means What to do
A “docs: dependency audit notes” pull request or a wfx-docs branch The account that opened it is compromised, “even if the workflow stage did not run,” says Aikido Revoke that account’s tokens, OAuth grants and SSH keys, close the pull request, delete the branch
A commit since November 10 titled “ci: refresh build cache,” “chore: add ci helper,” “chore: add dependency audit workflow,” “ci: restore” or “ci: cleanup” The payload was pushed Check the Actions run numbers around it
A run number missing right after that commit The payload probably ran, and its run was deleted Replace every secret and cloud role its workflows could reach
You ran npm install on an affected commit before it was reverted The script ran on your computer Treat the computer as compromised, heise and SafeDep advise
digitalelvis/neuronai-studio 3.14.3 from Packagist The package carries the payload Treat any machine that installed or opened it as compromised, Aikido says
github.com/onuragtas/docker-env v1.0.278 or v1.0.280 The payload is only in a workflow file, which Go builds never run Remove it anyway, Aikido recommends
A Fly.io app deployed from GitHub Its GitHub token may have been served publicly Reconnect GitHub and check the app, below

Where most of the tokens came from is unknown. On 173 probes, from 19 accounts, GitHub records the Fly.io GitHub App as the route in, StepSecurity found. Fly.io says its deploys from GitHub left a token in the build directory’s .git/config file, which some apps then served on the public internet, and that “Fly.io’s systems were not breached.” Aikido found that accounts hit by October’s GhostAction campaign “were significantly more likely than baseline to be impacted by wfx-loot,” though it believes different people ran the two.

How to check

  1. Find the probes. Run these for each account and organization you can write to, with the GitHub CLI. The second also finds the November 15 title.

    gh search prs --head wfx-docs --owner YOUR-ACCOUNT-OR-ORG
    gh search prs "dependency audit notes" --match title --owner YOUR-ACCOUNT-OR-ORG
  2. Find the payload commits. In each repository with a probe, look for the commit messages Aikido lists:

    git fetch --all
    git log --all --since=2026-11-10 --oneline --grep="ci: refresh build cache" --grep="chore: add ci helper" --grep="chore: add dependency audit workflow" --grep="ci: restore" --grep="ci: cleanup"
  3. Look for a missing run number. For each commit you find, open the repository’s Actions tab and look at the run numbers right after it. A gap is “the most reliable sign that the stealer ran,” StepSecurity says. Finding no wfx-loot artifacts or cache entries “does not mean data was not taken,” Aikido warns.

  4. Check your lockfiles for the Packagist and Go versions in the table. SafeDep also lists github.com/vkuznet/gocurl and github.com/onuragtas/redock.

  5. If you deploy to Fly.io from GitHub, open https://<your-app>.fly.dev/.git/config. If it returns a file, Fly.io says to add .git to your .dockerignore, redeploy, and “rotate any credentials that were ever committed to that repository.” Then sign in to Fly.io with GitHub again, since it is revoking every token its GitHub App issued. On November 13 it added that “not all tokens were actually revoked”: GitHub support told it the process “had started and then stopped partway through,” and GitHub would resume it.

If the payload ran

Aikido’s rule is blunt: “If the payload ran in your repository, assume every secret and cloud role available to its workflows is compromised.” The researchers’ steps, together:

  1. Lock the account. Revoke the compromised account’s personal access tokens, GitHub CLI tokens, OAuth grants and SSH keys, and turn on two-factor authentication.
  2. Replace every secret the job could read: repository, environment and organization secrets, npm and PyPI publishing tokens, and any long-lived cloud keys. StepSecurity says to check for package versions published after the attacker’s commit.
  3. Remove the cloud backdoors, which outlast new secrets. On AWS, look for role trust policies with "Principal": {"AWS": "*"} and access keys created since November 10, and in CloudTrail for session names starting with wfx-. On Azure, delete federated identity credentials named wfx-ci-*. On Google Cloud, review GenerateAccessToken and SignJwt calls through your GitHub workload identity pool.
  4. Re-run required checks on affected commits, since the script can forge passing ones, Chainguard notes.
  5. Clean up. Delete the wfx-docs and chore/ branches, close the probes, and remove any injected workflow steps, .ci-*.sh script, postinstall hook or build-audit-*.yml file.

Replacing AI keys

Deleting a GitHub secret doesn’t cancel the key in it. “If you delete a secret, the credential it held stays valid,” Anthropic’s GitHub Actions guide warns. Cancel each key where it was issued, then make a new one:

Secret Where it comes from How to cancel it
ANTHROPIC_API_KEY The Claude Console On the Console’s API keys page, open the key’s menu and choose Delete API key, says Anthropic
CLAUDE_CODE_OAUTH_TOKEN claude setup-token, a one-year token on your Claude plan On claude.ai, Settings > Account, log out of all sessions, which covers “tokens created with setup-token,” says Anthropic. It doesn’t affect API keys
OPENAI_API_KEY The OpenAI platform OpenAI says to “rotate your key immediately from the API Keys page” and to check its Usage page for use that isn’t yours
Gemini, Hugging Face, Cursor, Copilot and other AI keys Each provider In that provider’s settings

Putting the new key back into GitHub is a job many people hand to a coding agent. Don’t paste the key into the conversation: anything you type to an agent goes to your AI provider and is saved with the conversation, as our guide to keeping keys safe explains. In Enjoy, paste it and choose Keep as secret, then ask the agent to set it, for example with gh secret set ANTHROPIC_API_KEY --body "$ANTHROPIC_API_KEY". Enjoy runs the command with the key and shows the agent the output with the value hidden (how). That keeps the value out of the conversation; it doesn’t limit what the agent can do with it.

If the script ran on your own computer, through npm install, replace your coding agents’ logins too, not just the keys in GitHub. Our LiteLLM memo lists where Claude Code, Codex and GitHub CLI logins are cancelled.

What no one has said yet

  • Where most of the tokens came from. Only the Fly.io share is explained.
  • What GitHub says. “GitHub has not yet commented,” heise reported on November 13, and GitHub’s changelog has no entry on the campaign.
  • Whether Fly.io’s revocation finished. Its last update, on November 13, said GitHub would resume it.
  • How far it reached in private repositories, which the public counts can’t include, and whether the November 15 probe is the same attacker.

We’ll update this memo when GitHub or the researchers publish more.

8.Questions.

What is the “docs: dependency audit notes” pull request in my repository?

It’s the wfx-loot campaign’s probe, opened with a stolen token to test whether it can write to the repository. On its own it does nothing, StepSecurity says, but the account that opened it is compromised. Revoke that account’s tokens, OAuth grants and SSH keys, close the pull request, delete the wfx-docs branch, and check the repository for the attacker’s commits.

How do I know if wfx-loot stole my secrets?

Look for a missing workflow run number right after one of the attacker’s commits since November 10, such as “ci: refresh build cache” or “chore: add ci helper.” The script deletes its own run and logs, so Aikido calls a missing run number “the more reliable sign that the payload ran.” Finding no other trace doesn’t prove nothing was taken.

Were my OpenAI or Anthropic API keys stolen?

If the payload ran in a repository whose workflows could read them, assume so. The script takes every secret the job can reach and looks for credentials for OpenAI, Anthropic, Gemini, Cursor, Copilot, Codex and other AI tools. Deleting the GitHub secret isn’t enough: delete the key in the Claude Console or on OpenAI’s API keys page, and make a new one.

Was GitHub hacked?

No one has said so. StepSecurity says the attacker “does not exploit a bug in GitHub Actions” and uses valid tokens for real accounts. Where most of them came from is unknown. On 173 probes, they were tokens from Fly.io’s GitHub App, which Fly.io says some customers’ apps were serving publicly in .git/config files. GitHub hadn’t commented by November 13.

I deploy to Fly.io from GitHub. What should I do?

Sign in to Fly.io with GitHub again, since Fly.io began revoking every token its GitHub App had issued. Then open your app’s address followed by /.git/config: if a file comes back, Fly.io says to add .git to .dockerignore, redeploy, and replace any credential ever committed to that repository. Also check the repositories the app can reach for activity since November 10.

Is wfx-loot still going?

The researchers’ reports cover November 10 to 12, but GitHub’s search shows a probe from a wfx-docs branch opened on November 15 under a new title: “docs: add dependency audit notes.” Search for the branch and for “dependency audit notes,” not only the original title.

Enjoy Agentic Industries · Checked against its sources on