Skip to content

Security: tobsh/guardian-pattern-llm

Security

SECURITY.md

Security Policy

Scope

This repository is a proof of concept, not a production system. Use it as a reference implementation or a starting point for your own deployment — don't run it as-is for real users.

Reporting a vulnerability

If you find a security issue, please do not open a public GitHub issue. Instead, use GitHub's private vulnerability reporting:

Report a vulnerability

Please include:

  • A description of the issue and its impact
  • Steps to reproduce (minimal example preferred)
  • Which commit or branch you tested against
  • Any mitigation ideas you already have

I'll acknowledge the report within a few days and aim to ship a fix or mitigation as time allows. As a solo-maintained POC, there are no formal SLAs.

What counts as in-scope

  • Secrets or credentials leaked in the repo or build artifacts
  • IAM policies in the CDK stacks that grant more access than the service needs
  • Prompt-injection or jailbreak techniques that bypass the Guardian classifier and get the coach to give regulated advice
  • Bugs in the constitution rendering / verdict schema that break the fail-closed contract
  • Cognito / OAuth / API Gateway misconfigurations that expose the private /turn endpoints

Out of scope

  • Typos in German prompts
  • Model hallucinations that are not regulated advice (this is a POC, not a safety-verified product)
  • Cost overruns from your own deployment (tune the Lambda timeouts, memory, and Bedrock model choice for your workload)
  • Bedrock model capacity or rate limits

Rotating credentials

If you deployed this demo and suspect the AWS role or Cognito user pool was exposed, rotate:

  1. Destroy the stack: cd infrastructure && npx cdk destroy --all
  2. Rotate any AWS access keys you used for the deploy: AWS Console → IAM → Users → Security credentials
  3. Redeploy from a clean commit

The demo stores no user data in durable storage beyond CloudWatch logs — there's nothing to exfiltrate after teardown.

There aren't any published security advisories