Understudy is under active development. Security updates are applied to the latest
release branch on main.
| Version | Supported |
|---|---|
| 0.1.x | ✅ |
| < 0.1 | ❌ |
Understudy is designed around a zero-trust execution model for AI-generated remediation actions:
- Closed Action Enum: The LLM cannot execute arbitrary shell commands or code in
production. It may only synthesize typed plans conforming to
contracts/models.py. - Safety Kernel Verification (Z3): Every candidate plan that wins tournament scoring must be formally verified against system invariants before production actuation is permitted (Invariant K10).
- Triple-Layer Twin Egress Isolation:
- Kubernetes
NetworkPolicydenies egress from twin namespaces (except to the mirror gateway). - CoreDNS rewrite rules direct all outbound traffic to an egress stub.
- Workload environment variables point sensitive external endpoints to mock sinks.
- Kubernetes
- Read-Only Production Boundary: Fleet controllers and twin workers have read-only
access to
ust-prod. Only the orchestrator'sactuatenode holds actuation permissions. - No Secret Ingestion: Secrets and credentials are never stored in run records, playbooks, logs, or commit history.
We take the security of Understudy seriously. If you discover a security vulnerability or an isolation bypass, please report it privately rather than opening a public issue.
- Open a GitHub Security Advisory under the repository's "Security" tab, or
- Email the maintainers at
security@understudy.devwith:- A description of the vulnerability and its potential impact.
- Step-by-step reproduction instructions or a proof-of-concept.
- Any proposed remediation or mitigation.
- Initial Acknowledgement: Within 48 hours.
- Triage & Assessment: Within 5 business days.
- Resolution & Disclosure: Coordinated release and public advisory once a fix is verified.