GitHub Secret Stealer Attack: 4,837 Hijacked Accounts Planted Malware in 61,000 Repos

Thousands of developers woke up last week to find their GitHub accounts had been used to attack other people’s code. Starting on Tuesday, 10 November 2026, an unknown attacker used stolen login credentials for 4,837 GitHub accounts to open test pull requests in 61,764 repositories, then planted a secret-stealing script in thousands of them. Researchers call the campaign wfx-loot, and by scale it is one of the largest supply-chain attacks GitHub has ever seen.

The attack did not exploit a bug in GitHub itself. It exploited something simpler and harder to fix: trusted accounts with broad access. Once the attacker held a developer’s credentials, every repository that developer could write to became a target. The stolen secrets include cloud credentials and AI service keys, so the damage can reach into a victim’s Amazon, Google and Microsoft cloud accounts too.

What happened: 4,837 accounts, 61,764 repositories

The campaign began at about 12:45 UTC on Tuesday, 10 November 2026. Aikido Security, the security company that found it, disclosed the attack on 12 November. According to Aikido, the attacker opened probe pull requests from 4,837 compromised accounts across 61,764 repositories, and the secret stealer actually ran in at least 1,806 of them. Aikido warns that figure may understate the damage.

A separate analysis by SafeDep counted 8,093 public repositories that received malicious commits between 12:46 UTC on 10 November and 04:36 UTC on 12 November. SafeDep cautions that the true total is higher, because its search covered only public default branches.

The numbers make this campaign substantially larger than GhostAction, a similar wave of attacks in October 2026 that used about 1,900 hijacked accounts against roughly 30,000 repositories. The two campaigns are also connected: accounts hit in October were roughly 1,000 times more likely than other active accounts to be reused in this campaign, and in more than 100 repositories the new stealer was added to workflow files that still contained October’s older exfiltration step. Aikido has been careful not to blame the same actor, saying it does not attribute the November campaign to the group behind GhostAction at this stage.

How the wfx-loot attack worked

Developer working on a laptop covered in GitHub stickers
Photo: Christina Morillo / Pexels

The attack followed a consistent pattern: probe, payload delivery, revert and cleanup, and exfiltration. Each stage was designed to be fast, quiet and hard to trace.

The probe: a harmless-looking pull request

Each attack started with a pull request titled “docs: dependency audit notes”, opened from a branch named “wfx-docs”. The pull request changed nothing important. Its only purpose was to test whether the stolen account could write to the repository. These test pull requests were opened in 61,764 repositories, and a GitHub search later found 61,839 pull requests with that exact title.

Aikido warns that any account that opened one of the attacker’s test pull requests is compromised, even if the stealer never ran. The probe alone proves the attacker held that account’s credentials.

The payload: a 1,500-line secret stealer

Where the account could write, the attacker pushed the stealer into the repository’s CI workflows. The script is roughly 1,500 lines of shell code, and it arrived in one of three ways. A commit titled “ci: refresh build cache” added it to a workflow the repository already had. A commit titled “chore: add ci helper” added a script file and a postinstall hook in package.json, so the stealer also runs on any developer’s laptop that installs the package. A commit titled “chore: add dependency audit workflow” added a brand-new workflow on a short-lived branch.

Once a workflow ran, the stealer collected credentials from every source available on the runner: GitHub Actions secrets and variables, the workflow’s GITHUB_TOKEN, OIDC and Actions runtime tokens, AI service keys, and credentials stored in the memory of the runner’s own processes. It also swept up credential files for AWS, Azure, Google Cloud, Docker, Kubernetes, npm, PyPI and SSH, plus .env files and private keys.

The cover-up: reverts and deleted logs

Within about a minute of the payload running, the attacker reverted the malicious commit and deleted the workflow run that executed it. The most reliable sign the stealer ran is often a gap in the workflow run numbers right after the attacker’s commit.

What the stealer took, and where it went

The Hacker News and other researchers have published the indicators: the stealer script’s SHA-256 hash is 89695abb392d5cfd4972672961e644bb46baa98aa1ed7173c2a2816c762735cf, and the known exfiltration addresses are 194.150.166.244, 18.143.168.96 and 153.75.82.159, all on port 8080.

The stealer also reached into package ecosystems. Version 3.14.3 of the Packagist package digitalelvis/neuronai-studio was found to be malicious, and researchers advise treating any machine that installed or opened it as compromised; users should block that version and pin 3.14.2 or earlier. A Go module, github.com/vkuznet/gocurl, was also found carrying the stealer, as were versions 1.0.278 and 1.0.280 of the onuragtas/docker-env Go module.

The hidden danger: cloud backdoors that survive secret rotation

The most dangerous part of wfx-loot is not what it stole. It is what it may have left behind. The stealer carries code that swaps GitHub OIDC tokens for sessions in AWS, Google Cloud and Azure, and plants persistent backdoors in those cloud accounts. In Microsoft Entra ID, that means federated credentials with names starting “wfx-ci-“. In AWS, it means roles assumed through AssumeRoleWithWebIdentity, followed by the creation of new access keys.

No victim has yet confirmed that these backdoors were actually used in a real account. But the code is there, and that matters, because rotating secrets does not remove a backdoor. That is why researchers stress hunting for persistence changes in cloud audit logs, not just rotating credentials and moving on.

Part of a bigger pattern

Wfx-loot is the third major supply-chain attack to hit developers in a matter of weeks. The OPTIMJACK attack poisoned Optimizely scripts and served malware through the websites of 178 organisations, including CNN, the BBC and Zoom. Days earlier, the Fireworks AI breach exposed customer data belonging to Uber, Notion and Cursor. In each case, attackers broke into a trusted supplier, tool or account, and rode that trust past the target’s defences.

For GitHub users, the lesson is specific. A developer’s GitHub token is now a high-value credential in its own right, because one stolen token exposes every repository its owner can write to. Aikido counted 10,431 probe pull requests in repositories the compromised accounts did not own, reached through collaborator or organisation access. AI provider keys are a primary target too, because stolen model access can be resold, a practice known as LLMjacking.

How to check if your repositories were hit

Colourful programming code on a computer screen
Photo: Myburgh Roux / Pexels

If you maintain repositories on GitHub, or belong to an organisation that does, check now. Search every repository you can write to for the attacker’s fingerprints, going back to 10 November 2026.

First, search for probe pull requests titled “docs: dependency audit notes” and branches named “wfx-docs”, short-lived “chore/” branches, or “ci-check-” branches. Second, search your git history for the payload commits: “ci: refresh build cache”, “chore: add ci helper” and “chore: add dependency audit workflow”. Third, look for gaps in workflow run numbers around those commits, since the attacker deleted the runs that executed the stealer. Fourth, block digitalelvis/neuronai-studio version 3.14.3 and the affected Go module versions. Finally, search cloud audit logs: in AWS CloudTrail, look for AssumeRoleWithWebIdentity calls with a session name starting “wfx-“; in Entra ID, look for federated credentials starting “wfx-ci-“.

What to do if you were hit

Treat every secret, token and cloud role those workflows could reach as stolen. For every account that authored a probe pull request or a payload commit, revoke sessions, personal access tokens, SSH keys and OAuth app grants, then reset the password and enable phishing-resistant multi-factor authentication. Do the same for bot accounts, which are easy to overlook. Close the probe pull requests, delete the attacker’s branches, and confirm the revert really removed the payload.

Then rotate everything the workflows could read: repository and organisation secrets, cloud credentials, AI service keys, npm and PyPI tokens, and credential files on affected runners. Hunt your CloudTrail and Entra audit logs for the persistence changes above, and remove any federated credentials or access keys the attacker created. Treat any machine that installed the poisoned neuronai-studio package as compromised. If a probe pull request exists but you find no payload, rotate the account’s credentials anyway, because the probe alone means the account is compromised.

How to protect your GitHub account going forward

The cheapest defences are the ones GitHub already provides. Enable two-factor authentication with a passkey, which is phishing-resistant and needs no memorisation; GitHub’s own guidance recommends passkeys over codes. Use fine-grained personal access tokens with the minimum scopes each job needs, set expiry dates on them, and never use a broad classic token where a narrow one will do. Turn on push protection and secret scanning, and review your authorised OAuth apps, SSH keys and deploy keys regularly, revoking anything unfamiliar. And remember that stolen credentials usually start with a tricked human: learn how to spot a phishing email, because a single convincing message is still the easiest way to steal a developer’s login.

Wfx-loot shows how much damage one stolen login can do when it carries broad write access across dozens of repositories. Check your repositories today, rotate anything the attacker could have touched, and tighten the credentials before the next wave arrives.

Leave a Comment