Abridge — agentic threat model

Not certain from the listing — likely utilizes specialized automatic speech recognition (ASR) and clinical LLMs. Primary threats include clinical hallucinations, adversarial audio injection, and model bias leading to inaccurate medical summaries.

Not certain from the listing — processes highly sensitive audio recordings and patient-clinician dialogue. Primary threats include unauthorized retention of PHI, lack of secure data lineage, and data exfiltration of patient records.

Not certain from the listing — likely uses a structured pipeline (audio processing to EHR-compatible note generation) rather than an open-ended agentic framework. A key threat is indirect prompt injection, where a patient's spoken words maliciously manipulate the generated clinical note.

Not certain from the listing — requires highly secure, HIPAA-compliant cloud hosting with strict network isolation. Threats include container compromise or misconfigured cloud storage exposing sensitive clinical audio files.

Not certain from the listing — requires rigorous clinical validation, drift monitoring, and audit logging of clinician edits. Threats include silent degradation of transcription accuracy over time and insufficient logging of note modifications.

As an 'enterprise-grade' clinical tool, it must strictly align with HIPAA, HITECH, and SOC 2 Type II standards, integrating with EHR identity providers. Threats include unauthorized access to PHI, privilege escalation within the EHR, and regulatory compliance failures.

Not certain from the listing — likely operates as a downstream integration within EHR marketplaces (e.g., Epic App Orchard). Threats include API vulnerabilities in EHR integrations and cascading failures if the EHR schema changes.