VAREK Information

VAREK — deterministic, pre-execution verification of AI agent actions. VAREK is a pre-execution authorization runtime for autonomous AI agents. Before an agent acts, VAREK decides whether the action is provably authorized — returning SATISFIED, UNSATISFIED, or UNKNOWN from a decision procedure and enforcing the verdict at the kernel boundary with seccomp-BPF and seccomp user-notify. Anything not provably authorized doesn't run. Now shipping v1.34.0: VAREK on Kubernetes — a node agent and a signed Helm chart, with policies as a VarekPolicy resource. The agent's pod passes Pod Security "restricted", and 24-hour runs passed on Amazon EKS and OpenShift. Since VAREK Enterprise reached AWS Marketplace (v1.23), the runtime has added hosts by name, opt-in wildcards and an egress proxy that decides each request (v1.24–v1.26); program launches decided and enforced by Landlock (v1.27); credential release, so the agent holds only placeholders (v1.28); varek run with nothing set up first (v1.29); MCP servers confined and MCP calls decided, with OCSF events (v1.31); and an official signed container image for Docker and Podman, rootless (v1.32). Built on everything since v1.9: decided connections, a default-deny seccomp-BPF allowlist, decisions on the exact object delivered, an SMT decision procedure whose every authorization is checked by an independent certificate checker, and a hash-chained, signed verdict stream. Patent-pending (three USPTO provisional filings), MIT-licensed, open source. Our thesis comes from medicine: you don't deploy a system that's only usually right. VAREK holds agent autonomy to a fail-closed standard — formal proof, not probabilistic risk scoring. See VAREK in action: https://youtu.be/0PDDDeXYsr4 On AWS Marketplace: https://aws.amazon.com/marketplace/pp/prodview-6fdmjpuimvx64 varek-lang.org · github.com/kwdoug63/varek · MIT