
Safe loopback-only PoC and technical analysis for CVE-2026-116000, a server-side request forgery (CWE-918) in PrivateGPT's URI artifact loader, with reproduction scripts and remediation guidance.
English | Português do Brasil
Safe, loopback-only proof of concept for server-side request forgery in Zylon PrivateGPT's URI artifact loader. The issue was researched and reported by Fernando Viana (capivara-research).
| Item | Detail |
|---|---|
| Component | URI Artifact Loader |
| Function | load_file_from_uri() / _load_file_from_url() |
| File | private_gpt/server/ingest/uri_loader.py |
| Weakness | CWE-918 — Server-Side Request Forgery |
| Official affected versions | 1.0.0 and 1.0.1 |
| Official severity | CVSS 4.0 6.9 / CVSS 3.1 7.3 |
| Tested commit | 65b9b2b6b880a607ebf53cf4f2ac023ff9f5a0e0 |
| Vendor response in CVE record | No response / no documented fixed version |
VulDB reserved the identifier on 2026-10-24 and published it on 2026-10-25.
The canonical record credits fernando viana (VulDB User) as reporter and
lists versions 1.0.0 and 1.0.1 as affected. The exact official vectors and
record dates are preserved in
evidence/official-record-summary.md.
The CVSS 4.0 and 3.1 scores differ because the specifications model impact and supplemental metrics differently; they describe the same vulnerability.
An SSRF bug turns the vulnerable server into an HTTP client for the attacker. Imagine giving a courier an address without checking whether it points to a public office or a locked internal room. The request originates from the server, so it can reach destinations that the remote user may not reach directly.
The security question is not only "is this a valid URL?" It is also "is the server allowed to connect to this resolved address, including after every redirect?"
attacker-controlled artifact URI
-> load_file_from_uri()
-> _is_url() accepts any URL with scheme and network location
-> _load_file_from_url()
-> requests.get(..., allow_redirects=True)
-> server connects to the selected destination
The loader checks URL syntax but does not enforce a destination policy. It does not reject loopback, private, link-local, or otherwise sensitive address ranges. Redirects are enabled and are not revalidated against a destination allowlist.
The same dispatcher also falls back to _load_file_from_disk() for values that are not S3, URL, or Base64. That behavior is related to the public arbitrary-file-read report, but this repository's executable PoC focuses only on the CVE's SSRF classification.
The PoC starts a controlled server on 127.0.0.1, imports the real vulnerable module, and requests a fixed marker through load_file_from_uri(). It rejects the need for external infrastructure and does not probe cloud metadata.
The validation has two independent layers:
source-check.py confirms that the affected dispatcher and HTTP call exist without a private-address filter.reproduce.py proves actual server-side network access by receiving a unique marker from a loopback service.The workstation contained disclosure-time materials. Adapted, secret-free copies are preserved under artifacts/:
original-research-poc.py — adapted copy of the SSRF plus local-file-read laboratory script, with fixed loopback port 18766.VULDB-SUBMISSION.txt — curated submission material and the explicit limitations recorded during the research.Use poc/reproduce.py for routine validation. It imports the complete target module, selects a free port automatically, emits JSON, and has deterministic exit codes.
python3 -m pip install requests
python3 poc/source-check.py /path/to/private-gpt
python3 poc/reproduce.py /path/to/private-gpt
A vulnerable result exits with code 1 so the script can act as a negative security regression test:
"marker_received": true
[VULNERABLE] attacker-controlled URI reached the loopback HTTP service
Exit codes:
| Code | Meaning |
|---|---|
1 | Vulnerable behavior reproduced |
0 | Behavior not reproduced |
| other | Laboratory/setup error |
docker build -t cve-2026-116000 .
docker run --rm cve-2026-116000
The proven primitive allows the application worker to issue HTTP requests to an attacker-selected destination. Depending on network reachability, this can expose internal services or trigger state-changing internal requests. This repository does not claim access to any particular metadata service or secret.
Defenders should review application and egress logs for PrivateGPT workers connecting to loopback, RFC 1918, link-local, or otherwise unexpected address ranges shortly after ingest requests. DNS answers and redirect chains should be retained during incident review because filtering only the original URL is insufficient.
.
├── README.md
├── README.pt-BR.md
├── Dockerfile
├── artifacts/
│ ├── original-research-poc.py
│ └── VULDB-SUBMISSION.txt
├── poc/
│ ├── reproduce.py
│ └── source-check.py
├── evidence/
│ ├── validated-results.md
│ └── official-record-summary.md
├── SECURITY.md
└── LICENSE
| Date | Event |
|---|---|
| 2026-09-21 | VulDB submission #992813 recorded |
| 2026-10-24 | CVE-2026-116000 reserved |
| 2026-10-25 | CVE record published |
| 2026-10-26 | Safe PoC and validated evidence published in this repository |
Use this material only for authorized security testing, defensive validation, and education.