Skip to content

Security: hkevin01/secure-llm-assistant

Security

.github/SECURITY.md

Security Policy

Supported Versions

This is an internal, air-gapped deployment. All versions in active production use are supported. Reach out to the platform team for your deployment's current version.

Reporting a Vulnerability

Do not open a public GitHub issue for security vulnerabilities.

Report security issues to the security team via your organization's internal secure channel (e.g., security@company.internal or the InfoSec ticketing system). Include:

  1. Description of the vulnerability
  2. Affected component (guardrails, auth, translation pipeline, infra, etc.)
  3. Steps to reproduce or proof-of-concept (sanitized — no real credentials)
  4. Potential impact assessment

Response SLA:

  • Critical (RCE, auth bypass, credential leak): 24 hours acknowledgement, 72 hours patch
  • High (guardrail bypass, privilege escalation): 48 hours acknowledgement, 1 week patch
  • Medium/Low: Next sprint

Security Architecture Summary

Control Implementation Location
Authentication RS256 JWT / OIDC src/core/auth.py
Authorization RBAC role→permission src/core/auth.py
Prompt injection defence Regex guardrail src/guardrails/input_guard.py
Credential leak prevention Input + output regex redaction src/guardrails/
Network isolation K8s NetworkPolicy (no internet egress) infra/k8s/network-policy.yaml
Audit trail Append-only JSONL, metadata only src/core/logging.py
LLM determinism temperature=0 enforced src/core/llm_client.py
Hardened system prompt Anti-override instructions src/core/llm_client.py
Static code analysis AST-based SAST pre-LLM src/tools/python_analyzer.py

Known Limitations

  • The LLM system prompt does not guarantee injection-proof behaviour against all novel attack vectors; the input guardrail is the primary defence.
  • Translation output is a scaffold requiring human review — it must not be deployed to production without engineer validation.
  • The regex-based secret detection may produce false positives on code that resembles credential patterns.

There aren't any published security advisories