This repository is teaching material. It ships a deliberately vulnerable agent configuration
behind feature flags (STOCKROOM_WEAKNESSES=injection_unguarded, naive_retry, ...) and a policy
document in data/policy_docs/ that contains a seeded prompt injection. Those are features of
the Day 3 and Day 4 labs, not vulnerabilities to report. The shipped default has every weakness
fixed, and the red-team suite in promptfooconfig.yaml is part of the CI gate.
- Ways to make the default (all-fixed) harness execute an injected instruction, leak the system
prompt, exceed
MAX_STEPS/TOKEN_BUDGET, or call a tool with arguments that bypass validation. - Problems in the GitHub Actions workflows or the IAM documents in
docs/aws/that could expose credentials, widen the OIDC trust, or let a pull request trigger Bedrock spend. - Secrets or personal data accidentally committed to the repository.
Please use GitHub's private vulnerability reporting for this repository (Security → Report a vulnerability) rather than a public issue, so the details stay private until a fix is published. You should hear back within seven days.
If you cannot use that form, email the maintainer listed on the Genial Labs organisation page.
Only the main branch is maintained. Fixes land there and are tagged in the next release.