The vulnerability arises from the PyannoteDiarizationModelManager._load_fn function in speaches, which accepts an external model argument that is used to instantiate classes or execute code. This use of untrusted data aligns with CWE-470 and allows an attacker to supply a value that causes the loader to create a malicious class or execute arbitrary code. The flaw can lead directly to remote code execution, compromising confidentiality, integrity, and availability of the affected system.

The issue affects the speaches-ai speaches Python library, version 0.9.0‑rc.3 and earlier. The vulnerable component resides in src/speaches/executors/pyannote_diarization.py. Any environment that imports and uses the Pyannote Diarization model loading feature is at risk. The CPE indicates the product is speaches‑ai speaches. No later versions are documented in the release notes.

The CVSS base score is 6.9, a medium severity level. The EPSS score is absent, but a public proof‑of‑concept has been released, demonstrating that the flaw is actively exploitable over the network. The flaw is not yet listed in the CISA KEV catalog. An attacker can trigger the vulnerability remotely by sending a crafted model name or identifier to the _load_fn. Because the library has not yet released a fix or responded to issue reports, the window of vulnerability remains open, making the risk substantial for deployments that rely on that feature.