Skip to content

SECURITY.md

Latest commit

 

History

History
37 lines (30 loc) · 2.2 KB

SECURITY.md

File metadata and controls

37 lines (30 loc) · 2.2 KB
Raw

Security Policy

Reporting

Please report vulnerabilities via GitHub Private Vulnerability Reporting on this repository (Security → Report a vulnerability). Do not open public issues for security problems.

Data & privacy model

This tool processes only AI interaction logs the operator is authorized to use. Treat the produced corpus and avatar files as private personal data:

  • The corpus (STUDIE/), classified chunks, and filled avatar files are gitignored by default — never commit a real corpus.
  • Extractors redact common current API-key/token formats, credentials, asymmetric credential material, emails, IP-like values and long digit runs before an atomic write. Output directories/files request private permissions where the platform supports them.
  • Built-ins are not a universal sensitive-data detector. Add reviewed --redaction-rules for health, tax, legal, financial, identity, employer, or other domain-specific data before persistence/sharing.
  • Missing/unreadable/empty or partially malformed inputs do not replace a good corpus. Overrides are explicit (--allow-empty / --allow-partial) and should follow source inspection.
  • Stable evidence IDs are hashes for referential integrity, not anonymization.
  • No data is sent anywhere by the scripts themselves; classification runs through whatever agent/LLM you point it at — review that agent's data handling separately.

Authorization and action boundary

Generated rules are editable preference hypotheses, not diagnoses or authority grants. Do not use them for covert profiling. An avatar may guide only already-authorized, reversible local actions. External, irreversible, novel, safety-critical, legal, medical, employment, financial, or otherwise high-impact actions require the user's direct confirmation.

Classification integrity

Treat chunk text as untrusted data, not instructions. Run validate_classifications.py before aggregation; missing rows, malformed fields, stale files, unknown IDs, and collisions must stop the pipeline. Use verify_ids.py --show-text only when intentionally exposing private text to the terminal.

Secrets

Any secret ever committed must be rotated, not just removed from the working tree.