All articles

NewsNovember 9, 20265 min read

wfx-loot: 4,837 stolen GitHub accounts used to drain CI secrets

wfx-loot: using stolen GitHub tokens, the attacker probed 61,764 repositories and stole CI secrets and cloud credentials from over 1,800. What to check.

Hands holding a cloud with a padlock, surrounded by digital icons

Since 10 November 2026 a campaign has been running against GitHub that Aikido researchers have named wfx-loot, after the attacker’s own labels (“wfx” for the malware, “loot” for what it takes). It does not exploit a vulnerability in GitHub Actions: it uses valid, previously stolen tokens belonging to developers and bot accounts to push into CI pipelines a script that harvests secrets and cloud credentials. Help Net Security, SecurityWeek, BleepingComputer and The Hacker News have covered it, citing analyses by Aikido, StepSecurity and SafeDep.

The numbers

According to Aikido, the activity started around 12:45 UTC on 10 November:

  • 4,837 compromised accounts opened probe pull requests in 61,764 repositories;
  • 10,431 of them landed in repositories the account did not own, reached as a collaborator or organisation member: a single stolen token exposes every project its owner can write to;
  • 92% of the probe pull requests are still open;
  • the script ran in more than 1,800 repositories (BleepingComputer reports at least 1,835 cases).

SafeDep, quoted by SecurityWeek, counts 8,093 public repositories that received malicious commits, a figure that does not necessarily mean successful execution. About 50 compromised accounts can publish packages to npm, PyPI, Go, crates.io or Packagist.

The attack chain

  1. Permission check. With the stolen token the attacker opens a pull request titled docs: dependency audit notes, which does nothing by itself: it only tests for write access (T1078, valid accounts).
  2. Injection. Where possible, it adds a shell script of about 1,500 lines to an existing workflow, to a new workflow on a throwaway branch, or as an npm install hook, which also hits developers who install dependencies while the change is present.
  3. Collection. Actions secrets and variables, GITHUB_TOKEN, OIDC and runtime tokens; AWS, Azure and Google Cloud credentials, including those exposed by the instance metadata service on self-hosted runners (T1552.005); Git, SSH, Docker, Kubernetes, npm, PyPI, Vault and Terraform credentials; AI service keys; .env files, private keys, shell histories, recent commits. SecurityWeek adds that it reads the runner process memory.
  4. Cloud escalation and persistence. It exchanges OIDC tokens for AWS, Azure and GCP sessions, creates long-lived AWS keys (T1098.001), rewrites an AWS role’s trust policy to accept any account and adds federated identity credentials to Azure applications, which remain valid after tokens are rotated. It also forges passing status checks to get around branch protection.
  5. Exfiltration and clean-up. Tar archives are sent via HTTP POST to one of two attacker servers (T1041). Within a minute the change is reverted and the workflow run, logs included, is deleted (T1070): often the only trace is a skipped number in the run history.

The script also has settings for private repositories, so it may be active where researchers cannot see.

Packages carrying the payload

Automated publishing carried the payload into the Packagist package digitalelvis/neuronai-studio 3.14.3 and the Go module github.com/onuragtas/docker-env v1.0.278 and v1.0.280; in the Go tags the script sits only in a workflow file, which Go builds never run. Aikido says any machine that installed or opened the Packagist package should be treated as compromised.

Aikido sees overlaps with GhostAction, a campaign discovered in 2025 that returned in October 2026: accounts hit back then were more likely to be hit by wfx-loot too, and in more than 100 repositories the script was injected into GhostAction’s own workflow files. The two operations, however, are probably run by different actors.

On the October wave, Infosecurity Magazine reports Socket’s analysis: on 8 October two accounts with write access to hundreds of repositories added the workflow .github/workflows/security-audit.yml (commit “Add security audit workflow”), which also searches the Git history for credentials and sends them over HTTP to 193.32.204[.]199. It is worth looking for these indicators as well.

What to do now

  • Search for the indicators across all repositories: wfx-docs branch, the probe pull request, commits “ci: refresh build cache”, “chore: add ci helper” or “chore: add dependency audit workflow”, gaps in workflow run numbers and deleted runs in the audit log.
  • If the script ran, treat every secret and cloud role reachable by the workflows as compromised: revoke the tokens of the accounts involved and rotate secrets, cloud keys, registry tokens and API keys. Missing artifacts prove nothing: the data went straight to the attacker’s servers.
  • In the cloud, rotation is not enough: in every account tied to the affected workflows check IAM role trust policies, recent AWS users and keys, federated credentials on Azure applications and GCP service account usage.
  • Developer workstations: anyone who ran npm install while the malicious hook was present, or installed the Packagist package above, should be treated as compromised.
  • Reduce the attack surface: read-only GITHUB_TOKEN by default, production secrets in environments with required approval, OIDC conditions restricted to specific repositories and branches, mandatory reviews on protected branches, personal tokens with an expiry, restricted outbound traffic from self-hosted runners.

Impact for European organisations

There is no patch to install: the risk concerns anyone with developers or suppliers who have write access to repositories. For NIS2 entities, supply chain and development security is one of the required measures; if the compromise affects in-scope services, assess the notification duty (early warning to the national CSIRT within 24 hours; in Italy, CSIRT Italia).

How ARXDOME helps

On self-hosted runners and build servers with the ARXDOME agent, ARX Sentry collects events, processes and ports, and brings firewall logs, enriched with threat lists, into the SOC pages. With ARX Horizon you can look up public indicators, such as the address used in the October campaign, in the threat intelligence sources.

Last updated:

Any questions?

Let’s talk about your company’s needs, with no obligation.

Contact us

Related articles