CVE-2026-33625
8.8 · HighCode execution in LMDeploy through eval() of a model’s quant_dtype
A crafted quantization_config in a Hugging Face model made LMDeploy pass quant_dtype to eval(), so loading the model ran the attacker’s Python code.
Domain
Inference servers, agent memory and ML libraries.
6 write-ups, 6 vulnerabilities
CVE-2026-33625
8.8 · HighA crafted quantization_config in a Hugging Face model made LMDeploy pass quant_dtype to eval(), so loading the model ran the attacker’s Python code.
CVE-2026-77262
8.6 · HighMCP Atlassian opened any file_path passed to its Confluence upload tool, so a caller could read host files and push them out as attachments.
GHSA-wg9g-w2j2-8pgr
7.8 · HighMONAI’s NumpyReader loaded .npy and .npz files with allow_pickle=True, so a poisoned file in a medical imaging dataset ran code on load.
CVE-2026-34940
8.7 · HighA user allowed to create KubeAI Model resources could run shell commands inside model pods by adding metacharacters to the model URL.
CVE-2026-33980
8.3 · HighThree metadata tools of adx-mcp-server pasted table_name into KQL, letting an agent run arbitrary queries without going through execute_query.
CVE-2026-32247
8.1 · HighGraphiti pasted search filter node labels into raw Cypher, allowing a caller to bypass group filtering; writes depend on the database route and permissions.
| Identifier | Product | Vulnerability | CVSS | Published |
|---|---|---|---|---|
| CVE-2026-33625 | LMDeploy | eval() of quant_dtype from the model quantization_config | 8.8 High | |
| CVE-2026-77262 | MCP Atlassian | Path traversal in confluence_upload_attachment | 8.6 High | |
| GHSA-wg9g-w2j2-8pgr | MONAI | Unsafe deserialisation in NumpyReader | 7.8 High | |
| CVE-2026-34940 | KubeAI | Command injection through the Model URL in the Ollama startup probe | 8.7 High | |
| CVE-2026-33980 | Azure Data Explorer MCP Server | KQL injection through table_name in three metadata tools | 8.3 High | |
| CVE-2026-32247 | Graphiti | Cypher injection through SearchFilters.node_labels | 8.1 High |