Database/AI/ML frameworks & serving
NVIDIA Triton (Python backend shared memory): Out-of-bounds write in the Python backend
Impact
Out-of-bounds write in the Python backend
Who can reach it
Unauthenticated network
What to do
Patch; the Python backend's shared-memory region is the pivot in the Wiz chain
Fleet impact
How widespread
Very common - Triton is the default multi-model serving layer inside NIM and in many neocloud managed-inference products
Cost to remediate
daemon-restart (upgrade to 25.07, restart the serving process) - no host reboot, but it is a customer-visible endpoint outage
Why it hits the whole fleet
Unauthenticated remote chain: an oversized request leaks the internal shared-memory key, then an OOB write in the Python backend yields full RCE as the Triton process - i.e. theft of every model and every inference request on that server
References
Related entries
- LibreChat: agent Actions have no destination restrictions by default, reaching internal services via SSRFCVE-2025-69222 · LibreChat (agent Actions feature, outbound request allowlist)High
- picklescan: Misses `idlelib.pyshell.ModifiedInterpreter.runcode` gadgetCVE-2025-71340 · picklescanHigh
- picklescan: Misses `idlelib.run.Executive.runcode` gadgetCVE-2025-71342 · picklescanHigh
- JupyterLab: pasted cell keeps metadata.trusted, running script in the authenticated originCVE-2026-102831 · JupyterLab / Jupyter Notebook / JupyterLite (system-clipboard cell paste)High
- Keras (archive extraction utils): Path traversal in `keras/src/utils/file_utils.py`CVE-2026-11816 · Keras (archive extraction utils)High
- llama-server (KV cache state restore): Heap buffer overflow in `state_read_data`CVE-2026-43629 · llama-server (KV cache state restore)High
This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.