Skip to content

Security Policy

Supported Versions

Version Supported Notes
0.4.x :white_check_mark: Current developer preview (pin exact version)
0.3.x :x: Superseded
1.0.x :x: Retracted track (brief 2026-05 line; project returned to 0.x)
< 0.3 :x: Unsupported

Reporting a Vulnerability

If you discover a security vulnerability in ASHA, please report it responsibly.

Do not open a public GitHub issue for security vulnerabilities.

Instead, email: ajayrajan727@gmail.com

Include:

  • A description of the vulnerability
  • Steps to reproduce
  • Potential impact
  • Suggested fix (if any)

We aim to acknowledge reports within 48 hours and provide a status update within 7 days.

Scope

This policy covers:

  • PII detection and masking bypasses
  • Prompt injection / jailbreak evasion in ASHA's security layer
  • Fail-safe behavior failures (data leakage on error paths)
  • Dependency vulnerabilities in core asha package dependencies

Out of scope:

  • Vulnerabilities in third-party LLM providers (OpenAI, Anthropic, etc.)
  • Applications that disable privacy features (privacy=False, mode="off")
  • ML model behavior when using optional [ml] / [hardened] extras
  • ANCHOR framework-adapter edge cases outside CI stub coverage

Security Design Principles

ASHA follows these defaults:

  • Privacy-first: PII masking enabled by default
  • Fail-safe: Returns original or sanitized content on errors - never crashes the host app
  • No hidden downloads: Rule-based / lite path requires no model downloads
  • Local processing: Core security runs locally; no telemetry sent by default
  • Fail-closed on uncertain detections when hardened paths are available (REVIEW / restrictive defaults rather than silent ALLOW)

Best Practices for Users

  • Install with pip install asha-si; pin a tested version (e.g. asha-si==0.4.2) for production pilots
  • Use mode="strict" for sensitive workloads
  • Prefer wrap_llm() over auto_patch()
  • Run pip audit regularly on your environment
  • Do not log raw prompts containing PII in production
  • Validate ASHA on your own data before compliance use
  • Treat ANCHOR and experimental features as preview until 1.0
  • Never commit .env or detector weights to the public repository