Gitea: Permanent Fork PR Workflow Approval Gate Bypass

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:

 1// services/actions/notifier_helper.go:423-433
 2if count, err := db.Count[actions_model.ActionRun](ctx, actions_model.FindRunOptions{
 3    RepoID:        repo.ID,
 4    TriggerUserID: user.ID,
 5    Approved:      true,
 6}); err != nil {
 7    return false, fmt.Errorf("CountRuns: %w", err)
 8} else if count > 0 {
 9    log.Trace("do not need approval because user %d has been approved before", user.ID)
10    return false, nil
11}

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.

Primary

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

 1func ifNeedApproval(ctx context.Context, run *actions_model.ActionRun,
 2                    repo *repo_model.Repository, user *user_model.User) (bool, error) {
 3    // 1. don't need approval if it's not a fork PR
 4    // 2. don't need approval if the event is `pull_request_target` since the
 5    //    workflow will run in the context of base branch
 6    if !run.IsForkPullRequest ||
 7       run.TriggerEvent == actions_module.GithubEventPullRequestTarget {
 8        return false, nil
 9    }
10 
11    // always need approval if the user is restricted
12    if user.IsRestricted {
13        return true, nil
14    }
15 
16    // don't need approval if the user can write
17    if perm, err := access_model.GetDoerRepoPermission(ctx, repo, user); err != nil {
18        return false, fmt.Errorf("GetDoerRepoPermission: %w", err)
19    } else if perm.CanWrite(unit_model.TypeActions) {
20        return false, nil
21    }
22 
23    // ===== VULNERABLE BLOCK ==================================================
24    // don't need approval if the user has been approved before
25    if count, err := db.Count[actions_model.ActionRun](ctx, actions_model.FindRunOptions{
26        RepoID:        repo.ID,
27        TriggerUserID: user.ID,
28        Approved:      true,
29    }); err != nil {
30        return false, fmt.Errorf("CountRuns: %w", err)
31    } else if count > 0 {
32        return false, nil   // <-- permanent, unscoped bypass
33    }
34    // =========================================================================
35 
36    return true, nil
37}

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:
  text복사

   1if opts.Approved {
   2    cond = cond.And(builder.Gt{"`action_run`.approved_by": 0})
   3}

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

   1run.NeedApproval = false
   2run.ApprovedBy = doer.ID
   3if err := actions_model.UpdateRun(ctx, run, "need_approval", "approved_by"); err != nil {

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

   1NeedApproval bool                // may need approval if it's a fork pull request
   2ApprovedBy   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.

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.

Final action_run table contents (queried directly from SQLite)

 1id  triggered_by_uid  need_approval  approved_by_uid  status_text  ref               sha
 2--  ----------------  -------------  ---------------  -----------  ----------------  --------
 35   3                 0              1                Success      refs/pull/1/head  463c9114
 46   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)

 12026-05-23T21:38:53Z local-demo-runner(version:v1.0.6) received task 2 of job test, be triggered by event: pull_request
 22026-05-23T21:38:53Z workflow prepared
 32026-05-23T21:38:53Z Run Main pwn
 4...
 52026-05-23T21:38:53Z env:
 62026-05-23T21:38:53Z   BASE_TOKEN: ***
 72026-05-23T21:38:53Z ===== BYPASS DEMO - this workflow ran WITHOUT approval =====
 82026-05-23T21:38:53Z running as: uid=501(prakharporwal) gid=20(staff) groups=20(staff),12(everyone),...
 92026-05-23T21:38:53Z hostname:   PRAKHARs-MacBook-Air.local
102026-05-23T21:38:53Z uname:      Darwin PRAKHARs-MacBook-Air.local 25.5.0 Darwin Kernel Version 25.5.0...
112026-05-23T21:38:53Z GITHUB_TOKEN length = 40
122026-05-23T21:38:53Z outbound network test:
132026-05-23T21:38:53Z   HTTP 200
142026-05-23T21:38:53Z repo contents:
152026-05-23T21:38:53Z total 0
162026-05-23T21:38:53Z drwxr-xr-x  2 prakharporwal  staff   64 24 May 03:08 .
172026-05-23T21:38:53Z drwxr-xr-x  5 prakharporwal  staff  160 24 May 03:08 ..
182026-05-23T21:38:53Z ===== END DEMO =====
192026-05-23T21:38:53Z 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.

The logic embodies a misreading of the unit of trust.

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:

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.

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:
   text복사

    1.gitea/workflows/ci.yml

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

    1name: CI
    2on: [pull_request]
    3jobs:
    4  test:
    5    runs-on: [ubuntu-latest]
    6    steps:
    7      - name: hello
    8        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):
   bash복사

    1name: CI
    2on: [pull_request]
    3jobs:
    4  test:
    5    runs-on: [ubuntu-latest]
    6    steps:
    7      - name: pwn
    8        env:
    9          BASE_TOKEN: ${{ github.token }}
   10        run: |
   11          echo "===== BYPASS DEMO - this workflow ran WITHOUT approval ====="
   12          echo "running as: $(id 2>/dev/null || echo n/a)"
   13          echo "hostname:   $(hostname 2>/dev/null || echo n/a)"
   14          echo "uname:      $(uname -a 2>/dev/null || echo n/a)"
   15          echo "GITHUB_TOKEN length = ${#BASE_TOKEN}"
   16          curl -fsS --max-time 5 https://example.com/ -o /dev/null -w "egress HTTP %{http_code}\n" || echo "(no network)"
   17          ls -la
   18          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 document into the Gitea web editor and inadvertently keep any leading whitespace before name:, Gitea will reject the file with yaml: line 2: mapping values are not allowed in this context. No action_run row is created in that case (the workflow detector silently skips unparseable files), and the bypass will look like it didn't fire when in fact the code path was never reached. name: must sit at column 0.

Open a database shell on the Gitea host. The action_run table contains the smoking gun.

SQLite (default)

 1sqlite3 /var/lib/gitea/data/gitea.db

 1SELECT
 2  ar.id,
 3  ar.trigger_user_id,
 4  ar.need_approval,
 5  ar.approved_by,
 6  CASE ar.status
 7    WHEN 1 THEN 'Success' WHEN 2 THEN 'Failure' WHEN 3 THEN 'Cancelled'
 8    WHEN 4 THEN 'Skipped' WHEN 5 THEN 'Waiting' WHEN 6 THEN 'Running'
 9    WHEN 7 THEN 'Blocked' ELSE CAST(ar.status AS TEXT) END AS status,
10  ar.ref
11FROM action_run ar
12JOIN repository r ON r.id = ar.repo_id
13WHERE r.owner_name = 'maintainer' AND r.name = 'gate-bypass-poc'
14ORDER BY ar.id;

Expected output after Phase 6:

 1id  trigger_user_id  need_approval  approved_by  status   ref
 2--  ---------------  -------------  -----------  -------  ----------------
 3N   <attacker uid>   0              <maint uid>  Success  refs/pull/1/head
 4N+1 <attacker uid>   0              0            Success  refs/pull/2/head

The N+1 row demonstrates the bypass: same trigger user, never approved (approved_by = 0), yet need_approval = 0 and executed successfully. This row is schema-indistinguishable from a normal trusted-user run; an audit pipeline scanning the table cannot tell the bypass apart from a legitimate write-permission user's run.

PostgreSQL / MySQL

The same query, against database_name, with \c gitea / USE gitea; first.

Three options, in order of preference.

Option A (recommended) — Scope approval to the commit, not the user

Extend FindRunOptions with a CommitSHA predicate (it already exists as the field at models/actions/run_list.go:70) and reframe the gate check as "has the current PR head SHA been approved by this user".

 1--- a/services/actions/notifier_helper.go
 2+++ b/services/actions/notifier_helper.go
 3@@
 4-    // don't need approval if the user has been approved before
 5-    if count, err := db.Count[actions_model.ActionRun](ctx, actions_model.FindRunOptions{
 6-        RepoID:        repo.ID,
 7-        TriggerUserID: user.ID,
 8-        Approved:      true,
 9-    }); err != nil {
10-        return false, fmt.Errorf("CountRuns: %w", err)
11-    } else if count > 0 {
12-        return false, nil
13-    }
14+    // don't need approval if a prior run for this same PR-head commit was approved
15+    if run.CommitSHA != "" {
16+        if count, err := db.Count[actions_model.ActionRun](ctx, actions_model.FindRunOptions{
17+            RepoID:        repo.ID,
18+            TriggerUserID: user.ID,
19+            CommitSHA:     run.CommitSHA,
20+            Approved:      true,
21+        }); err != nil {
22+            return false, fmt.Errorf("CountRuns: %w", err)
23+        } else if count > 0 {
24+            return false, nil
25+        }
26+    }

This preserves the legitimate UX intent ("don't re-prompt on the same PR after a transient runner failure causes a re-run") without granting trust to future PRs or new commits on the same branch. Every new commit on the PR — including a force-push that replaces the workflow file — would re-trigger the gate, exactly as a security-sensitive deployment expects.

Option B (stricter) — Require a merged contribution, like GitHub

Replace the "approved before" branch with "has this user had a PR merged in this repo before":

 1if merged, err := issues_model.HasUserMergedPullRequestInRepo(ctx, repo.ID, user.ID); err != nil {
 2    return false, fmt.Errorf("HasUserMergedPullRequestInRepo: %w", err)
 3} else if merged {
 4    return false, nil
 5}

This matches the typical OSS-maintainer mental model: a contributor whose PR has been merged is trusted enough to run CI without further approval. Approving a workflow run is not merging code; the current logic conflates the two.