NLTK EUVD-2026-66396

Exploitation requires the target Python application to pass attacker-controlled strings as regex patterns directly to NLTK's Text.findall() or TokenSearcher.findall() methods without prior sanitization or timeout enforcement. Applications that build regex patterns internally from fixed templates, or that restrict user input to plain keyword strings before constructing patterns, are not exploitable. The CVSS 4.0 vector (PR:N/UI:N/AV:N) reflects deployment where these methods are reachable over a network without authentication - if an authentication layer gates access to the search endpoint, PR:L or PR:H would apply, materially reducing the attack surface. Exploitation is further limited by the GIL-serialized nature of CPython: the DoS affects only the targeted process instance, not other processes or the host OS. Multi-process application servers (e.g., gunicorn with multiple worker processes) limit blast radius to a single worker per attack request unless the attacker submits requests concurrently.

The CVSS 4.0 score of 8.7 (AV:N/AC:L/AT:N/PR:N/UI:N/VC:N/VI:N/VA:H/SC:N/SI:N/SA:N) reflects worst-case network-exposed deployment with no authentication barrier. Real-world risk is deployment-dependent: only applications that accept user-supplied regex strings and route them to NLTK's findall methods are exploitable. Applications that construct regex patterns internally from fixed templates are not affected. No EPSS data was provided in the input and no KEV listing exists, indicating no confirmed active exploitation at time of analysis. However, ReDoS is a trivially reproducible vulnerability class - any attacker who identifies an exposed NLTK findall endpoint can construct a proof-of-concept in minutes without specialized tooling. The availability-only impact (C:N/I:N/A:H) limits attacker objectives to disruption rather than data exfiltration, but the ability to continuously crash or freeze a shared Python process makes this a credible denial-of-service primitive. No conflicting signals were identified in the available data; the primary uncertainty is the fraction of NLTK deployments that expose user-supplied regex input.

An attacker identifies a web application that accepts a search query and passes it as a regex to NLTK's Text.findall() - for example, a corpus search tool or NLP API. The attacker submits a payload such as the pattern (a+)+b against a long string of 'a' characters, causing the Python re engine to explore exponentially many backtracking paths. The single-threaded GIL ensures the stuck regex blocks all other requests, effectively bringing the application to a standstill until the process is restarted. No authentication is required if the search endpoint is publicly accessible.