Wiz "Pwn Request Reckoning"
Tags: ops, supply-chain, ci-cd, github-actions, pwn-request, pull-request-target, detection-engineering First seen: 2026-10-26 (Wiz Research publication) Source: https://www.wiz.io/blog/github-actions-security-pwn-request-reckoning (Shay Berkovich, Oct 26, 2026, full text captured by this wiki) Related: Gheist page (Part-1 threat model this updates), GhostAction page, LiteLLM 1.107.0 page (a prod-approval-gate bypass on the same trigger family), Mini Shai-Hulud page (the May TanStack wave chained a pull_request_target Pwn Request with cache poisoning + OIDC extraction)
Summary
Wiz's Part-3 of its GitHub Actions research line updates the Pwn-request threat model for GitHub's 2026 platform changes. The default-case exploit (any fork PR triggering a pull_request_target workflow that checks out fork code) is now materially harder, and the residual risk has a specific, auditable shape: the protections live inside actions/checkout, so anything that fetches code another way is unprotected; the protections apply to two triggers, so other privileged triggers are unprotected; and script injection was never a checkout bug at all.
The 2026 hardening timeline (as published)
| Date | Change |
|---|---|
| 2026-06-18 | actions/checkout@v7.0.0 blocks fork-code checkouts by default on untrusted triggers (pull_request_target, workflow_run-on-PR) |
| 2026-06-26 | Read-only cache tokens at platform level for untrusted triggers (kills cache poisoning) |
| 2026-07-20 | Fork-checkout protection backported to all active majors: v2.8.0, v3.7.0, v4.4.0, v5.1.0, v6.1.0 (v1 excluded); mutable tags (@v4) inherit automatically |
| 2026-09-10 | Fine-grained cache controls (cache-mode) GA |
| 2026-09-17 | Workflow Execution Protections GA (event policies, actor rules); default event policy in evaluate mode |
| 2026-12-02 | GitHub begins ENFORCING the default event policy that blocks pull_request_target in public repos without an explicit event policy (private/internal repos NOT affected) |
The residual attack surface (durable findings)
- THE SHA-PINNING PARADOX. Supply-chain best practice, pinning actions to exact commit SHAs, can silently opt a workflow OUT of the platform protection, because the guard is logic inside
checkoutitself: a SHA-pinned ref that resolves to a pre-Jul-20 release (e.g.actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955=v4.3.0) runs legacy code with NO fork-checkout protection, while@v4(→ v4.4.x) or@v4.4.0are protected. Audit rule: every SHA-pinnedactions/checkoutref must resolve to ≥ v2.8.0 / v3.7.0 / v4.4.0 / v5.1.0 / v6.1.0 / v7.0.0. This is a NEW false-security inversion: the hardened-looking config is the vulnerable one. - PROTECTIONS ARE ACTION-LOCAL. Any fetch that isn't
actions/checkoutoperates unrestricted underpull_request_target: directgit clone+checkout $PR_SHA,gh pr checkout,gh repo clone, and third-party checkout actions configured with a PR head ref. - PROTECTIONS ARE TRIGGER-LOCAL (two triggers only).
pull_request_targetandworkflow_runare covered;issue_commentbot-command workflows (/deploy,/test,/cherry-pick) that check outrefs/pull/N/mergeare fully exploitable on checkout@v7 with no flag, note the inconsistency Wiz flags: GitHub's own cache controls classifyissue_commentas low-trust while checkout protections ignore it.repository_dispatchruns with secrets and is exploitable when payload data influences the checkout ref. Authorization must hold from event-fire to code-run: check the commenter's CURRENT permission (getCollaboratorPermissionLevel, write+, NOTauthor_association), pin the exact head SHA (the/mergeref is mutable, an attacker can push after the approving comment), and abort if the PR updated after the comment. - SCRIPT INJECTION IS UNAFFECTED BY EVERY 2026 CHANGE.
${{ }}interpolation of attacker-controlled event fields (pull_request.title, comment body, dispatch payload) intorun:blocks is code execution with the job's secrets and no checkout required. The fix is structural: pass untrusted context through environment variables (quoted), never into the script text. - THE DECEMBER DEFAULT IS A DEFAULT, NOT A BAN. Workflows that add an explicit permissive event policy, labelers, preview deploys, fork CI with secrets, are exactly the ones handling untrusted input with elevated privileges, and private/internal repos are outside the enforcement entirely.
Hunt / audit checklist (as captured)
- Repo-wide grep:
pull_request_target+ any explicit event policy; audit WHY the opt-in exists. actions/checkout@<40-hex>refs → resolve the SHA → flag any ref older than the Jul-20 backports.run:steps that fetch code outsideactions/checkouton untrusted triggers.issue_comment/repository_dispatchworkflows with secrets +refs/pull/.*/mergeor payload-derived refs.run:blocks interpolatinggithub.event.*fields directly (quote-and-env is the fix).- December-2 readiness: inventory public-repo workflows that will silently stop without an event policy, expect a wave of rushed permissive policies around the enforcement date; that wave is the new opt-in population to re-audit.
Why it matters (durable reads)
- Pwn requests remain the credential-to-CI path that fed this year's biggest supply-chain stories (the Mini Shai-Hulud TanStack wave chained one with cache poisoning and OIDC extraction; LiteLLM's prod approval gate fell to a second-account approval, same trust-boundary family). Platform hardening moves the class underground; it does not delete it.
- The SHA-pinning paradox inverts a decade of Actions advice, detection should key on the PINNED-OLD case specifically, because that population believes itself hardened.
- The enforcement-date effect (Dec-2) creates a predictable, dated defender event: audit who writes permissive event policies in November.