[Retest #3751918] Issue title injection still present in agent-triage.yml and agent-fix.yml

RETEST-3751918.md

@@ -0,0 +1,100 @@
+# Retest — HackerOne Report #3751918
+**Retested by:** @subchatsecure  
+**Date:** 2026-06-26  
+**Status: VULNERABILITY STILL PRESENT**
+
+---
+
+## What was checked
+
+Both flash-list agent workflows were pulled from `main` and compared against the original report.
+
+---
+
+## agent-triage.yml — Still Vulnerable
+
+**File:** `.github/workflows/agent-triage.yml`  
+**Commit fetched from:** `main` branch, 2026-06-26
+
+Line 39-40 (current `main`):
+```yaml
+prompt: |
+  You are running on CI (GitHub Actions).
+  Triage issue #${{ github.event.issue.number }}.
+  Title: ${{ github.event.issue.title }}    ← STILL INJECTED VERBATIM
+```
+
+The only change observed vs the original report is the addition of a proxy content-filter header:
+```yaml
+env:
+  ANTHROPIC_BASE_URL: https://proxy.shopify.ai/vendors/anthropic
+  ANTHROPIC_CUSTOM_HEADERS: |-
+    Shopify-Security-Scan: paranoid-path-template
+    Shopify-Security-Scan-Mode: block
+```
+
+**This is not a fix.** The root cause — attacker-controlled text interpolated directly into the agent prompt at the YAML level before the process starts — is unchanged. A server-side content filter is bypassable and does not address the injection at the source.
+
+`author_association` check: **still absent** on the `labeled` trigger path.
+
+---
+
+## agent-fix.yml — Still Vulnerable (No Proxy Filter Either)
+
+**File:** `.github/workflows/agent-fix.yml`  
+**Commit fetched from:** `main` branch, 2026-06-26
+
+Line 84-85 (current `main`):
+```yaml
+prompt: |
+  You are running on CI (GitHub Actions). There is no human to confirm with — act autonomously.
+  ...
+  YOUR ASSIGNED ISSUE: #${{ github.event.issue.number }}
+  TITLE: ${{ github.event.issue.title }}    ← STILL INJECTED VERBATIM
+```
+
+This workflow:
+- Has `contents: write`, `pull-requests: write`, `issues: write` permissions (unchanged)
+- Allows `Bash(git push -u *)`, `Bash(gh pr *)`, `Bash(GH_TOKEN=* gh pr *)` (unchanged)
+- Has `AGENT_PR_TOKEN: ${{ secrets.SHOPIFY_GH_ACCESS_TOKEN }}` (unchanged)
+- **Does NOT have the `Shopify-Security-Scan` proxy header** — not even the partial mitigation applied to agent-triage.yml
+- `author_association` check: **still absent** on the `labeled` trigger path
+
+---
+
+## What a real fix looks like
+
+The correct fix is to never pass attacker-controlled data as inline prompt text. Use a safe indirection:
+
+```yaml
+# FIXED agent-triage.yml
+- name: Triage issue
+  uses: anthropics/claude-code-action@...
+  env:
+    ISSUE_NUMBER: ${{ github.event.issue.number }}
+    # Do NOT pass title here — Claude reads it via gh issue view
+  with:
+    prompt: |
+      Triage issue #$ISSUE_NUMBER.
+      Use `gh issue view $ISSUE_NUMBER` to read the title and body.
+      Read .claude/skills/triage-issue/SKILL.md and follow its instructions.
+```
+
+And add an `author_association` guard on the `labeled` trigger:
+```yaml
+if: >-
+  github.event.label.name == 'agent-triage' &&
+  github.event.sender.type != 'Bot' &&
+  (
+    github.event.sender.login == github.repository_owner ||
+    contains(fromJSON('["OWNER","MEMBER","COLLABORATOR"]'), github.event.label.author_association)
+  )
+```
+
+---
+
+## Conclusion
+
+The reported vulnerability (#3751918) is **not fixed**. Both workflows continue to interpolate `${{ github.event.issue.title }}` directly into the Claude agent prompt with no escaping or env-var indirection. The only mitigation applied (`Shopify-Security-Scan` proxy header) is limited to `agent-triage.yml` and is server-side only — it does not remove the injection surface.
+
+`agent-fix.yml` — the highest-severity workflow with full write access — has received zero changes.