Listen to this Post
How the CVE Works
The vulnerability resides in the `preprocess_string()` function within Hugging Face Transformers’ testing utilities (v4.48.3). The function uses a flawed regular expression to parse docstrings containing code blocks. The regex pattern contains nested quantifiers (e.g., (\\n+)+), which, when processing malicious input with excessive newline characters (\\n), triggers catastrophic backtracking. This forces the CPU into inefficient computations, exponentially increasing processing time. Attackers can craft a payload with thousands of newlines, causing prolonged high CPU usage, rendering the application unresponsive—a classic ReDoS attack.
DailyCVE Form:
Platform: Hugging Face
Version: v4.48.3
Vulnerability: ReDoS
Severity: Moderate
Date: May 19, 2025
What Undercode Say:
Analytics:
- Impact: Application downtime, CPU exhaustion.
- Attack Vector: Malicious docstring payloads.
- Mitigation Rate: High (patch available).
Exploit Command:
from transformers.testing_utils import preprocess_string malicious_payload = "\n" 100000 Triggers ReDoS preprocess_string(malicious_payload) Exploit
Protection Code:
Patch: Replace vulnerable regex with non-backtracking pattern
import re
safe_pattern = re.compile(r"\n{1,100}") Limit quantifier
def safe_preprocess_string(text):
return safe_pattern.sub(" ", text)
Mitigation Steps:
1. Upgrade to Transformers >= v4.48.4.
2. Sanitize Inputs: Reject docstrings with >100 newlines.
- Regex Audit: Use tools like `regexlint` to detect unsafe patterns.
Detection Script:
Scan for vulnerable versions pip list | grep "transformers==4.48.3"
Workaround:
Disable `testing_utils` in production via:
import transformers transformers.testing_utils = None Disable module
References:
- GitHub Advisory: GHSA-xxxx-xxxx-xxxx
- NVD: CVE-2025-XXXX
- Patch Commit: huggingface/transformersxxxxx
Sources:
Reported By: github.com
Extra Source Hub:
Undercode
