domains / families / security / phi-prompt-guard

phi-prompt-guard

separate

Runtime, prompt-time behavioral guardrail that helps reduce PHI exposure in LLM-assisted workflows by detecting PHI-bearing prompts, avoiding unsafe tool actions that would pull more PHI in, and redirecting users toward de-identified or synthetic inputs. Use when the user is about to paste, query, or read clinical/patient data, or when an action (DB query, file read, tool output) may pull PHI into the conversation. Honors a [PHI-OK] attestation for synthetic / test data.

When to use

• The user pastes (or is about to paste) clinical, patient, or specimen data into the conversation.
• The agent is about to run a database client (`psql`, `mysql`, `mongo`, `duckdb`, `sqlite3`, `bq`, `snowsql`, `redis-cli`, `clickhouse-client`, `cqlsh`) or a dump tool (`pg_dump`, `mysqldump`, `mongodump`) against an environment that may co…
• The agent is about to read a file whose contents may contain identifiers (lab reports, EHR exports, accession-keyed CSVs).
• The user pastes a **clinical document by type** — H&P, SOAP note, discharge summary, progress note, op note, path report, radiology report, HL7 v2 message, FHIR resource JSON, CCD/CDA — these are PHI-by-construction even when no trigger ke…
• The user **attaches an image or screenshot** of an EHR, lab portal, chart, or clinical document. Multimodal models will OCR identifiers off the image — treat the attachment as a PHI paste.
• Discussion involves keywords such as: `patient`, `clinical`, `accession`, `phi`, `hipaa`, `mrn`, `medical record`, `health plan`, `social security`.
• The environment is unknown — assume production with PHI by default.

When not to use

• Preventing PHI already pasted from reaching the model
• Acting as a pre-submit filter or middleware

Skill page

Source