- Socket's guide says several 2026 incidents used valid credentials, so no input check could reject them. Its controls limit damage and speed revocation.
- In the LiteLLM incident, the attacker appeared as three accounts, which looked like three separate breaches. Socket says it was one compromise that began with a single exposed key.
- Count independently protected credentials, not people. Review who holds bypass rights, retire stored publish tokens, and rehearse revocation.
Three approvers can still be one lock
A release that needs three accounts looks safe. Socket's new guide to stolen-credential and insider threats says that comfort can be false. Its test is simple. How many separately protected things must an attacker break before a malicious release ships?
"Separately protected" is the phrase that matters. Counting people measures your org chart. Counting independent credentials measures your exposure.
What happened at LiteLLM
Socket describes an incident at LiteLLM on 21 October 2026. Branch protection blocked an unauthorized edit to the private release workflow. The edit had skipped the required pull request review and status checks.
Then a credential linked to a privileged developer account overrode the block. It took five seconds. The edit went through.
Socket says the rule was set up correctly and did its job. The account held bypass rights, and those rights won. Socket reads the five-second gap as a scripted, planned step.
On paper, three accounts acted. One pushed past the protection. One approved the release. One erased the trail.
Socket says the finished investigation tells a different story. There was one chain of compromise, not three. It began with an administrative key that had been committed in public. That key exposed the GitHub CLI token of an organization owner. The attacker used that access to plant workflows and take other engineers' tokens.
Socket adds that the owner's token could have lifted the approval rule altogether. Using three accounts seems to have helped the attacker stay hidden. It was not needed to get past the gate.
Why a valid credential changes the job
Part one of Socket's series dealt with attackers who hold no credential. Those defenses are largely predictable. If you pin an action to a commit SHA, a retagged action cannot run.
This second part is harder. The attacker holds a real credential. Nothing in the request looks malformed. GitHub, npm and the cloud provider see normal work.
Socket states the limit plainly. Its controls cannot stop every account takeover. They shrink the damage, add hurdles that are protected separately, and cut the time between the first bad act and revocation. Socket calls this containment and detection, not prevention.
Some gaps sit outside two-factor authentication. Socket points out that three kinds of credential skip the second factor: personal access tokens, SSH keys and deploy keys. They work from any location for as long as they live. A deploy key also belongs to the repository, not to a person. Removing a departing employee does not remove it.
How the attacks looked from inside
Socket cites the 8 September 2025 takeover of npm maintainer Qix. The attacker had registered a look-alike domain three days earlier. The victim received a message from that domain warning that their login protection needed attention. A copied sign-in page then collected both the password and the one-time code.
Once inside, the attacker published tainted releases of chalk, debug and more than a dozen related packages. Socket puts the combined weekly downloads at two to three billion. npm took the releases down in about ninety minutes. To the registry, each publish looked like the maintainer's own.
GhostAction worked differently. Socket says the attacker read a repository's workflows to learn the names of its secrets. Then the attacker pushed a new workflow that sent the values to a fixed address. One commit message posed as a security improvement.
GitGuardian counted 817 repositories and 3,325 stolen secrets in the first wave. It counted 772 repositories in a second wave through September 2026. Socket notes that figures differ between sources.
Branch protection and CODEOWNERS review apply to pull requests into the main branch. A direct push to any other branch gets no review. Socket says a workflow can run on that branch at once.
What Socket recommends
For publishing, Socket recommends trusted publishing. The registry trusts one named repository, workflow and environment. The stored token is then deleted. npm's version became generally available on 31 July 2025.
Setting it up does not revoke anything. Socket warns that old tokens stay valid until someone sweeps them out.
Socket also advises granting the publish permission only to a narrow job that runs no project code. Small teams cannot always split duties. For them, Socket suggests a staged release that a human approves with a security key. It also suggests alerts for odd configuration changes.
One date matters. On 2 December 2026, GitHub begins enforcing a default policy that blocks the pull_request_target trigger. It applies to public repositories without a policy of their own. Socket says to add an explicit policy first if you rely on that trigger.
What this changes for the people who sign off
This is the publication's view, not Socket's. An approval count reported to auditors or a board may overstate real separation. If one administrator can read the tokens of every approver, three sign-offs are one control wearing three names.
So the number to ask for is different. Ask for how many independent credentials an attacker must break, not how many people approve.
Someone must own that answer. Name one person accountable for the credential inventory, including bypass holders, deploy keys and stored publish tokens. Name another for the revocation drill. Without owners, both lists go stale.
Questions for your security lead
Ask who holds bypass rights on protected branches and rulesets. Ask whether one administrator can read the tokens of several approvers. Ask which publish tokens still sit in storage as secrets.
Ask whether an alert fires when a workflow file changes outside a merged pull request. Socket also flags the string toJSON(secrets) as a warning sign. No legitimate workflow serializes the whole set of secrets.
Then run the revocation drill, which Socket calls essential, and time it.
A gate is only as strong as the account allowed to open it.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Socket.





