DeepPavlov CVE-2026-112098

Exploitation requires the target to process an attacker-influenced DeepPavlov configuration through the Config Reference Resolver (the resolve() function in deeppavlov/core/common/params.py) in versions up to and including 1.7.0; the CVSS UI:P metric indicates passive victim interaction (loading/running a crafted config) is needed rather than fully autonomous remote exploitation. LIMITING FACTORS: the attacker must get malicious config content into the victim's pipeline, the victim must load it, and there is no evidence of exploitation at scale (no KEV, no EPSS) - so this is realistically an abuse of untrusted configuration input, not a drive-by remote RCE against a default network service.

Signals point to a LOW real-world priority despite the 'code injection' label. The only provided CVSS is a VulDB CVSS 4.0 base vector scoring just 2.1, driven by all-Low impacts (VC:L/VI:L/VA:L) and a passive user-interaction requirement (UI:P), with exploit maturity E:P (proof-of-concept). A public POC exists (github-poc exploit archive), but there is NO CISA KEV listing and NO EPSS score in the input, so there is no evidence of active or widespread exploitation - POC availability and active exploitation are distinct and only the former is confirmed here. CONFLICT TO NOTE: the vector declares AV:N (network) and PR:N (no privileges), yet the vulnerability is in a Config Reference Resolver (resolve() in params.py) and the vector also carries UI:P - this strongly implies the victim must load or process an attacker-supplied configuration, i.e. the true access vector is closer to local/config-dependent than genuine remote-unauthenticated. The sole data source is the VulDB CNA (cna@vuldb.com); there is no independent vendor advisory, and the vendor has not responded, so affected-version and impact data are single-sourced and unverified. CWE-74 (injection) combined with 'code injection' could in principle mean arbitrary code execution (far above a 2.1), so the Low impact rating itself may be understated - but the exploitation precondition (loading untrusted config) keeps realistic risk low.

An attacker distributes or convinces a developer/operator to load a crafted DeepPavlov configuration file; when the Config Reference Resolver's resolve() function processes the malicious reference, attacker-controlled code is injected and executed in the context of the DeepPavlov process. A public proof-of-concept (poc.zip) has been released, lowering the bar to reproduce the issue. Because CVSS carries UI:P, the attack hinges on the victim loading the malicious config rather than fully remote unauthenticated delivery.