AllenNLP EUVD-2026-109790

Exploitation requires the host application to pass attacker-controlled configuration (FromParams-style params containing a 'type' value) into AllenNLP's Registrable configuration machinery, specifically Registrable.resolve_class_name in allennlp/common/registrable.py; the attacker manipulates that Type argument to drive unsafe reflective instantiation. Limiting factors: only applications that load or deserialize untrusted AllenNLP configurations are exposed (the library alone, processing trusted configs, is not); the CVSS vector specifies UI:P, so a victim must process the malicious configuration; affected versions are AllenNLP up to and including 2.10.1; and because the product is retired and the vendor unresponsive, no fixed version is available - avoiding untrusted configurations or discontinuing use of the unsupported library is the only remaining mitigation stated by the circumstances.

The signals point to a low real-world priority. The supplied CVSS 4.0 base score is 2.1 (Low), with AV:N/AC:L/AT:N/PR:N but UI:P and only Low vulnerable-system impact (VC:L/VI:L/VA:L, no subsequent-system impact). That Low impact rating sits in tension with the CWE-94 (code injection) classification and the PoC title, which describe unsafe reflective instantiation leading to code execution - a code-injection flaw that actually yields arbitrary code execution would normally warrant High impact, so the vendor-assigned Low impact likely reflects that exploitation depends on the host application feeding attacker-controlled configuration into the library rather than a directly reachable remote entry point. KEV status is NO (not CISA-listed) and no EPSS score was provided, so widespread or automated exploitation is not evidenced. The vuln.today Decision label is 'monitor', the lowest urgency tier, and the disclosure originates from a single source (VulDB) with an unresponsive vendor. Net: this reads as a genuine but low-priority, config-dependent issue in a retired product, not a high-risk emergency; the primary limiting factor - the application must process untrusted AllenNLP configuration - keeps real-world exposure narrow.

An application built on AllenNLP accepts a model/experiment configuration (params with a 'type' field) from an untrusted source and passes it into the FromParams machinery. An attacker submits a crafted configuration that manipulates the Type argument so Registrable.resolve_class_name reflectively resolves and instantiates an attacker-chosen class, resulting in code injection. A public PoC reference exists (a GitHub writeup) but it is not verified as working exploit code, and the victim application must actively process the malicious configuration (UI required).