deepset-ai FARM CVE-2026-120334

Exploitation requires an application using deepset-ai FARM ≤0.8.0 to pass attacker-influenceable input into the vocab_filename argument of EmbeddingTokenizer.from_pretrained; the traversal occurs when that filename is used to locate the vocabulary file during tokenizer loading. Limiting factors: the CVSS vector specifies UI:P, so some passive user interaction is required; impact is confined to limited confidentiality disclosure (VC:L, no integrity or availability effect); and exploitability depends on the host application actually exposing vocab_filename to untrusted input rather than hardcoding it. Because the product is no longer maintained, no patch is available - the only complete remedy per the evidence is to stop using the unsupported library, as the vendor did not respond and issued no fix.

This is a genuinely low-priority issue, and the signals are mutually consistent rather than conflicting. The CVSS 4.0 vector (AV:N/AC:L/AT:N/PR:N/UI:P/VC:L, base 2.1) reflects a network-reachable, low-complexity path-traversal that requires passive user interaction (UI:P) and yields only limited confidentiality impact (VC:L) with no integrity or availability effect. EPSS is not supplied. This is NOT in CISA KEV - there is no confirmed active exploitation. The vuln.today decision label is 'monitor,' the lowest actionable tier, agreeing with the low base score. On exploit availability the data is weaker than it first appears: the description states 'the exploit has been published,' but the multi-source intelligence qualifies the PoC as a reference only, not verified as working exploit code - so a public PoC reference exists, but it is not verified as working exploit code. The most important mitigating context is that the product is end-of-life ('only affects products that are no longer supported by the maintainer') and the vendor did not respond, so no fix will be issued. Net assessment: a high-visibility advisory with low real-world risk - a limited-impact disclosure flaw in a deprecated niche ML library, not an emergency.

An application built on deepset-ai FARM ≤0.8.0 passes an externally-influenced value into EmbeddingTokenizer.from_pretrained's vocab_filename argument; an attacker supplies a traversal sequence so the loader reads a file outside the intended tokenizer directory, disclosing limited file contents. The CVSS vector implies this needs some passive user interaction (UI:P) but no authentication, and the low attack complexity means no special timing or grooming. A public PoC reference exists, but it is not verified as working exploit code.