What happened
Replaying the November wfx-loot wave
- 1
Tokens in served files
Not dated; used from Nov 11, 2026Fly.io GitHub App tokens sit in .git/config in the build directory of apps that serve the whole directory at their public URL, and they are read from there.
- 2
Write access
Nov 10 to 12, 2026In the November wave, stolen tokens from 4,837 accounts opened probes in 61,764 repositories. The actors changed workflows, install hooks and existing shell scripts where their permissions allowed it.
- 3
Secrets out
Nov 10 to 12, 2026Workflow-file variants requested the full secrets context available to the job. Install hooks and ordinary shell scripts collected credentials reachable by their process. Successful runs sent credentials to the actor.
- 4
Attempts on the clouds
Nov 10 to 12, 2026; no success confirmedStepSecurity found failed or empty cloud credential checks in the Nextcloud run. The stealer also carried code to establish persistence. No source confirmed a cloud session or persistence in a victim.
Impact
Aikido reported credential theft from at least 1,835 repositories during the November wave. StepSecurity reported no affected WebdriverIO package release. One Packagist version contained the payload and two Go module versions contained it only in workflow files. The stealer included code for cloud sessions and persistence. No source confirmed either in a victim.
Could this have been prevented?
The assessed controls could have limited the impact of this incident.
Incident control assessments
- 0
- Prevented
- 15
- Contained
- 7
- Not covered
How it could have been prevented
The same replay with the controls in place. Green marks where a step is refused, blue where it is bounded, and red where no control reaches it. Open a row to read the control, why it acts, and its caveat.
Replaying the November wfx-loot wave
1 · Tokens in served files
1 Not covered2 · Write access
3 Contained2 Not covered3 · Secrets out
5 Contained3 Not covered1 on other clouds4 · Attempts on the clouds
7 Contained1 Not covered2 on other cloudsKey facts
- GitGuardian found 80 precursor patches in 12 repositories from October 24 to 29. The November wave began about 12:45 UTC on November 10. By 15:00 UTC on November 11, Aikido counted 61,839 probe pull requests from 4,837 accounts in 61,764 repositories. StepSecurity counted more than 63,000 probes the following morning
- Personal access tokens carried most activity; bot account tokens, a GitHub CLI OAuth token one victim reported as leaked, and Fly.io GitHub App user tokens were also used, the last for 173 probe pull requests from 19 accounts, per StepSecurity. How most tokens were obtained is unconfirmed
- Fly.io said its GitHub App tokens, left in .git/config in the build directory, were read from apps that served that directory publicly; its own systems were not breached, and deploys with flyctl, Fly.io API tokens and running apps were not affected. It began revoking every user token of the app on November 12, and on November 13 said GitHub's revocation had stopped partway
- GitGuardian documented changes to existing workflows, new workflows on short-lived branches, npm install hooks and existing shell scripts. The seven November workflow-file runs still visible in its sample were held for approval. It also found successful uploads in 18 of 19 sampled repositories where the install hook ran
- At least 1,835 repositories ran the payload with credentials exfiltrated, per Aikido on November 12, and Endor Labs counted 8,926 repositories with payload evidence as of 21:35 UTC that day. Of 101 copies StepSecurity examined, 57 posted over HTTP only, 1 over HTTP and Telegram and 43 over Telegram only; fake passing checks were built in but disabled in all 101
- The stealer included code for cloud sessions and persistence. StepSecurity reported a successful upload from nextcloud/docker, where the logged cloud attempts failed or returned empty results. No source confirmed a cloud session or persistence in a victim
Sources
Every fact in this breakdown comes from the public reporting below. The replay is a model of that reporting, not a reconstruction of the environment it describes.
- 1.wfx-loot: A CI Secret Stealer Pushed With Stolen GitHub CredentialsGitGuardian · Nov 13, 2026 · ResearchOctober precursor, November approval holds, variant-specific secret exposure and observed uploads
- 2.wfx-loot campaign pushes CI secret stealer to tens of thousands of repositoriesAikido Security · Nov 12, 2026 · ResearchTimeline, counts, delivery variants, payload capabilities, cloud escalation capability, indicators
- 3.WFX Phantom: the GitHub Actions secret stealerStepSecurity · Nov 12, 2026 · ResearchToken types, delivery, cleanup, observed cloud attempts, exfiltration shift, victims
- 4.GitHub Actions supply chain attack targeting CI secrets at scaleEndor Labs · Nov 13, 2026 · ResearchPayload versions, counts, victims, access vector
- 5.Security: GitHub tokens exposed by apps deployed from GitHubFly.io · Nov 13, 2026 · Vendor reportToken exposure, scope, response, revocation status
- 6.Hackers compromise 4,800 GitHub accounts to steal cloud credentialsBleepingComputer · Nov 12, 2026 · NewsDisclosure date, campaign start, named affected organizations
The estate in the replay is illustrative and drawn at the level of attack stages; the victim's real environment is not public. Verdicts describe what a control does when it is in place before the incident. They are not an assessment of the victim's security program.