ENFORCEAUTH 2027 — NO. 04 OF 18 — SUPPLY CHAIN
By Mark Rogge, CEO of EnforceAuth | 6 min read | for platform, DevOps and AppSec engineers
For one hour and 44 minutes on October 21, 2026, pip install litellm handed you a credential stealer with a valid signature.
The timeline, from The Hacker News: the attacker held three sets of developer GitHub CLI credentials, including workflow permissions. One was an org-owner token. With it they disabled branch protection, then ran the release workflow against a lookalike repository so that the resulting pip and Docker artifacts were signed by the legitimate pipeline. The malicious pip release was live from 11:39 to 13:23 UTC; the Docker image from 11:46 to 14:20 UTC. The payload, "privcollect," swept system, browser, developer, cloud and AI secrets and sent them to a lookalike domain. On Linux and macOS, a sitecustomize.py hook fetched command-and-control instructions every hour.
About 4.5 hours later, Wiz observed the attacker validating AWS instance-role credentials harvested from CI runners and Kubernetes nodes. The suspected initial access, which LiteLLM has not confirmed, is that GitHub CLI OAuth tokens were recovered from logs in which an AI coding agent had embedded them in response bodies. It was LiteLLM's second supply-chain compromise of 2026, after the March incident attributed to TeamPCP.
Signing answers the wrong question
Every control that fired in this incident worked. The artifacts were signed. The signature verified. The signature said: this came from the LiteLLM release pipeline. It did. That was the problem.
Provenance tells you where an artifact came from. It does not tell you whether the sequence of actions that produced it was authorized: whether that token was allowed to disable branch protection, whether a release could run from that repository, whether any of it should happen without a second approver. Those are authorization questions, and nobody in the pipeline was asking them.
A signature is a statement of origin. Authority is a statement about whether the actions that produced the artifact were permitted. In 2027, your supply chain needs both, and it has one.
The counter-argument is that branch protection is the authorization control. It was, until a single token turned it off. A control that can be disabled by one credential it is supposed to govern is a setting, not a policy. The policy question is "who may disable branch protection on this repository, with what approval, during what window," and that question needs to be evaluated outside the thing being protected.
The pipeline is an identity, and it has a model in it
Here is what changed between the March compromise and the October one. In March, the attack was a classic credential theft. In October, the leading theory is that an AI coding agent wrote OAuth tokens into its own response bodies, those responses landed in logs, and the logs were the source.
That is not a GitHub problem or a LiteLLM problem. It is the 2027 problem in miniature. Agents in the development loop produce text that contains whatever they were given. Give them a token to do a task and the token is now in a transcript. Our design partners count 82 non-human identities per human, and the ones growing fastest are coding agents and the pipelines they feed. Each one is an identity with credentials, acting without a person present, producing logs that are themselves secret-bearing.
Forrester expects AI-generated bugs and regressions to cause global outages in 2027 as patch velocity outruns human review. The LiteLLM incident is the security mirror of that: release velocity outran authorization review, because there was no authorization review, only a signature.
What I expect in 2027
By the end of 2027, I expect the major CI/CD and package ecosystems to treat the pipeline as an autonomous identity subject to runtime authorization, not just to provenance attestation. Concretely:
- At least one widely used registry will require, not merely support, a policy attestation on privileged repository actions (branch-protection changes, workflow permission grants, force-pushes) in addition to artifact signatures.
- Secret scanning will extend to agent transcripts and tool-call logs as a default, because that is where the LiteLLM tokens were allegedly found.
- "Who could have authorized this release" becomes a standard incident-response question, and teams without a decision log will answer it by reading GitHub audit events for a week.
The control, stated as what ships
At EnforceAuth we express pipeline authority as OPA and Rego policy, evaluated at a policy decision point in front of the action:
- Who may disable branch protection on a release repository: named identities, a required second approver, a change window, and DENY otherwise. The org-owner token is still powerful. It is no longer sufficient alone.
- Where a release workflow may run. A workflow identity is bound to its repository. A lookalike repository has a different identity, and the signing step is an authorized action only for the real one.
- Agent logs as secret-bearing. Policy on what an agent may write to a transcript sink is enforceable. Whether tokens appear in response bodies is a model-behavior problem; whether that response is stored unredacted in a log an attacker can read is an authorization problem.
EvidenceIQ, part of the EnforceAuth platform, records each of these decisions tied to the policy version, keeps the relationships between the token, the identity, the repository and the artifact, and lets you replay the chain. After an incident like October 21, "which token authorized which step, and what was the state of branch protection at 11:39 UTC" is a query, not an archaeology project. Decision logs ship as signed Git bundles with one-to-four-year retention and export to your SIEM.
What we do not do: replace Sigstore, GitHub's audit log, or your artifact scanner. Keep those. They answer origin. We answer authority.
The pipeline did what it was told on October 21. In 2027, make sure someone checked whether it was allowed to.
EnforceAuth is the runtime authorization layer for AI agents and non-human identities, built on OPA and Rego. See enforceauth.com and read the EvidenceIQ announcement.
About EnforceAuth
EnforceAuth is the OPA/Rego authorization control plane. The customer's OPA decides what people, services, and AI agents may do at action time. Verdict covers high-impact agent actions. EvidenceIQ keeps the record behind every decision.

