Listen to this Post
The CVE-2025-XXXX vulnerability in Hugging Face Transformers stems from inefficient regex handling in the `DonutProcessor` class. The `token2json()` method uses the pattern <s_(.?)>, which is susceptible to catastrophic backtracking when processing maliciously crafted input strings. An attacker can exploit this by supplying a string with nested or unbalanced tags, causing exponential time complexity in regex evaluation. This leads to excessive CPU consumption, resulting in a Denial-of-Service (DoS) condition. The flaw affects versions ≤4.51.3, allowing disruption of document processing tasks relying on the Donut model.
DailyCVE Form:
Platform: Hugging Face Transformers
Version: ≤4.51.3
Vulnerability: ReDoS via `token2json()`
Severity: Moderate
Date: Jul 11, 2025
Prediction: Patch expected by Jul 25, 2025
What Undercode Say:
Check installed version: pip show transformers Exploit PoC (malicious input): curl -X POST --data "<s_<<<<<...>>>>>>" http://victim-api/process Mitigation (upgrade): pip install transformers>=4.52.1
How Exploit:
- Crafted input with nested `
` tags triggers regex backtracking. - Causes CPU spike, degrading service performance.
- Exploitable via API endpoints using
DonutProcessor.
Protection from this CVE:
- Upgrade to Transformers ≥4.52.1.
- Implement regex timeout mechanisms.
- Sanitize input before processing.
Impact:
- Service disruption due to CPU exhaustion.
- Affects document parsing workflows.
- Potential API downtime.
Sources:
Reported By: github.com
Extra Source Hub:
Undercode
