Technical impact
Crawl4AI from version 0.9.0 up to, but not including, 0.9.5 does not filter the CosineStrategy model_name field supplied in untrusted Docker API request bodies, allowing an authenticated client to select which HuggingFace model the server will load. An attacker can point the model_name to a repository that hosts a malicious pickle weight file, which under PyTorch versions lower than 2.6 can cause the server to execute arbitrary Python code. Even without code execution, a malicious model can exhaust disk space or memory, leading to denial of service.
Affected Systems
The vulnerability affects the Crawl4AI product released by unclecode. Any deployment running the 0.9.0 through 0.9.4 releases is impacted. Versions 0.9.5 and later have fixed the input validation issue on the CosineStrategy endpoint.
Risk and Exploitability
The CVSS score is 7.7, indicating a high potential for impact. EPSS data is not available, and the vulnerability is not listed in CISA’s KEV catalog. Exploitation requires that the attacker has legitimate credentials to the server, as the flaw is triggered by authenticated clients via the Docker API. Once authenticated, the attacker can supply a crafted request to load a malicious model, enabling remote code execution or resource exhaustion. The overall risk is moderate to high, contingent on the strength of the authentication mechanism in place.