See the signal.
Understand the risk.

01 / SOURCE HYGIENE

A credential-like literal in a release commit

If a real credential is committed, someone with repository access may be able to misuse it. A text match alone does not establish that a secret is live.

const api_key = "fictional_demo_value_not_a_real_key";

const apiKey = process.env.SERVICE_API_KEY;

What CodeLens observes

The static checker flags an assignment that resembles a hardcoded credential. The report identifies the file without publishing its value.

Remediation to review

Remove the literal, rotate any genuinely exposed key, review repository history and inject the replacement through a managed server secret.

What remains unverified

Removing a literal does not prove rotation, revoke historical access, or make a browser-delivered secret safe. Key custody and runtime permissions require separate verification.

Fixture check: the before snippet triggers credential_like_literal; the after snippet removes that specific signal. No broader “safe” verdict.

02 / CI TRUST BOUNDARIES

A privileged pull-request workflow worth reviewing

A privileged workflow that checks out untrusted changes can put repository credentials or release integrity at risk. Exploitability depends on the actual ref, permissions and subsequent steps.

on: pull_request_target
jobs:
  review:
    steps:
      - uses: actions/checkout@v4

on: pull_request
permissions:
  contents: read
jobs:
  review:
    steps:
      - uses: actions/checkout@v4
        with:
          persist-credentials: false

What CodeLens observes

The checker detects pull_request_target alongside checkout in a workflow and asks for human review. It does not execute the workflow or claim an exploit succeeded.

Remediation to review

For ordinary untrusted-PR checks, use a nonprivileged event, minimize token permissions, avoid exposing secrets and inspect what code each step runs.

What remains unverified

This is an illustrative pattern correction, not a complete runnable workflow or security approval. A privileged workflow can have legitimate uses with independently reviewed safeguards.

Fixture check: the before snippet triggers privileged_pull_request_workflow; the after snippet removes that specific signal. No broader “safe” verdict.

03 / AI EVIDENCE INTEGRITY

An AI finding points at code that is not there

An unsupported finding can send a developer toward an unnecessary or damaging change. An agent response must not become a verified claim merely because it arrived through MCP.

export const role = 'reader';

export const role = 'admin';

What CodeLens observes

The evidence validator binds the submission to a snapshot and checks its quoted line. A mismatched excerpt is labelled evidence_not_matched; the conclusion remains an unverified agent claim.

Remediation to review

Ask the agent to inspect the exact commit, correct the reference and explain the inference before a developer reviews any patch.

What remains unverified

Matching an excerpt is not proof that the interpretation is correct. Signed tokens do not replace independent commit provenance, authorization or human review.