Untrusted Model Loading in Crawl4AI by UncleCode
CVE-2026-119120

7.7HIGH

Key Information:

Vendor

Unclecode

Status
Crawl4ai
Vendor
CVE Published:
29 October 2026

What is CVE-2026-119120?

Crawl4AI versions up to 0.9.4 have a vulnerability allowing authenticated users to exploit the CosineStrategy model_name field in Docker API requests. This flaw fails to adequately sanitize input, permitting attackers to specify models from untrusted repositories, which may include malicious pickle weight files. Consequently, this can lead to arbitrary code execution under Python's PyTorch framework, particularly affecting systems below version 2.6, along with potential resource exhaustion impacting disk space and memory.

Affected Version(s)

crawl4ai 0.9.0 < 0.9.5

crawl4ai 0.9.5

References

https://github.com/unclecode/crawl4ai/security/advisories...
vendor-advisory
https://www.vulncheck.com/advisories/crawl4ai-0.9.0-befor...
third-party-advisory

CVSS V4

Score:
7.7
Severity:
HIGH
Confidentiality:
High
Integrity:
High
Availability:
High
Attack Vector:
Network
Attack Complexity:
High
Attack Required:
Physical
Privileges Required:
Undefined
User Interaction:
None

Timeline

  • Vulnerability published

  • Vulnerability Reserved

Credit

manus-use
ntohidi
.