Details
if ! grep -q 'PoC_FILTER_CONTROL=PASS' <<<"${OUTPUT}" || ! grep -q 'PoC_FILTER_ATTACK_SENT depth=20000' <<<"${OUTPUT}" || [[ "${STATE}" != "true 137 false" ]]; then
echo "[PoC] FAIL: nested filter did not produce the expected cgroup OOM termination" >&2
exit 1
fi
echo "PoC_FILTER_RECURSION=PASS depth=20000 OOMKilled=true ExitCode=137"
echo "[PoC] SUCCESS: an approximately 120 KB authenticated request terminated a 512 MiB API process"
Create `reproduction/frontend-placeholder.html`:
```html
<!doctype html><title>PoC backend reproduction placeholder</title>
From the directory containing these files, run:
chmod +x reproduction/run.sh reproduction/*.sh 2>/dev/null || true
./reproduction/run.sh
Expected: A low-privileged request below the server request-size ceiling terminates the API process through unbounded filter parsing.
Observed: Depth 1,000 returned HTTP 200 in 14 ms. The 120,044-byte depth-20,000 request disconnected, and Docker recorded OOMKilled=true, ExitCode=137. The run emitted PoC_FILTER_RECURSION=PASS.
Verification and controls: The client creates an ordinary account and project, asserts the depth-1,000 route control is HTTP 200, submits depth 20,000, and the host script accepts only the attack marker plus Docker OOMKilled=true and ExitCode=137.
Observed evidence:
- depth=1000 target_bytes=6044: HTTP 200
- depth=20000 target_bytes=120044: RemoteDisconnected
- target container: OOMKilled=true, ExitCode=137
- PoC_FILTER_RECURSION=PASS
Suggested remediation
Enforce the intended authorization, size, cardinality, recursion, or lifecycle boundary before the sensitive operation described above; fail closed; release partial resources on every exit path; and add a regression test that preserves the exploit and negative-control oracles.
Severity
CVSS v4.0: 7.1 (High) — Vector: CVSS:4.0/AV:N/AC:L/AT:N/PR:L/UI:N/VC:N/VI:N/VA:H/SC:N/SI:N/SA:N
This assessment is preliminary and pending maintainer confirmation. The score was recalculated with the FIRST CVSS v4.0 reference implementation on 28 August 2026.
Disclosure context and attribution
AI-assisted analysis helped surface this issue; the behavior was independently reproduced and validated in an isolated environment.
Reported by the University of Sydney security research team:
- Ziyue Wang (@Zyy0530)
- Liyi Zhou (@lzhou1110)
- Strick Sheng (@Str1ckl4nd)
- Maurice Ng (@mauriceng98)
- Chenchen Yu (@7thParkk)
We are happy to answer questions, provide additional verification details, or validate a candidate patch.