TL;DR
A researcher found six SGLang vulnerabilities, and the worst allow unauthenticated remote code execution. SGLang serves large language models such as DeepSeek and Qwen. CERT/CC published the advisory on July 30, 2026. No patch exists yet, and the maintainers have not responded.
- Total: 6 CVEs
- Severity: 3 Critical · 2 High · 1 Medium
- Actively exploited: None confirmed
- Highest severity: 9.8 (Critical · CVSSv3) — CVE-2026-15969
- Action: Apply the latest security updates now
Notable CVEs
| CVE | CVSS (CVSSv3) | Type | Fixed in | Status |
|---|---|---|---|---|
| CVE-2026-15969 | 9.8 | CWE-502 | — | Not exploited |
| CVE-2026-15971 | 9.8 | CWE-95 | — | Not exploited |
| CVE-2026-15976 | 9.8 | CWE-502 | — | Not exploited |
| CVE-2026-15977 | 7.5 | CWE-522 | — | Not exploited |
| CVE-2026-15978 | 7.5 | CWE-306 | — | Not exploited |
| CVE-2026-15974 | 6.5 | CWE-918 | — | Not exploited |
Why it matters
SGLang runs inference for popular open models. Many teams expose it as an API service. An unauthenticated flaw therefore hands attackers a direct path to the host. CERT/CC notes that “exploitation does not require authentication in most cases.” So any internet-facing instance faces serious risk.
How the attacks work
Three of the SGLang vulnerabilities lead to remote code execution. CVE-2026-15969 abuses a weak unpickling denylist to run commands through crafted payloads. CVE-2026-15976 triggers unsafe pickle deserialization when loading model weights. CVE-2026-15971 enables code execution through an optional dumper subsystem.
The other three expose data. CVE-2026-15974 is an SSRF and local file read that can reach “internal metadata, secrets, and services.” Next, CVE-2026-15977 leaks API keys through the server-info endpoint. Finally, CVE-2026-15978 lets a remote attacker exfiltrate entire model weights.
Affected versions
The flaws affect current SGLang builds where the vulnerable endpoints are reachable. CERT/CC lists no fixed release, because no patch is available.
Exploitation status and mitigation
No in-the-wild exploitation has been reported. However, full technical details are now public, which raises the risk. Since no fix exists, act on the workarounds in the CERT/CC advisory. First, restrict access to the service and keep it off untrusted networks. Also set the pickle IPC option to “false” in environ.py, and disable endpoints you do not use.
Support Our Threat Intelligence
If you find our CVE report and cybersecurity news helpful, consider supporting our work.