Sondera Presents Autoformalization Research for AI Agent Policy Control

Sondera announced in a press release that its research on converting natural language policies into formally verified controls for AI agents has been accepted at the ICML 2026 Agents in the Wild workshop and the FLoC 2026 LLM-Solve workshop. A related security tool, GolemHalt, will be demonstrated at Black Hat Arsenal.

The research paper, titled "Autoformalization of Agent Instructions into Policy-as-Code," describes a pipeline that reads organizational rules written in natural language and compiles them into formally verified Cedar policy code. Each rule is checked by a theorem prover and tested through adversarial simulations before deployment.

In benchmark testing using MedAgentBench, the system automatically formalized more policy rules than previous hand-coded approaches and successfully blocked all unsafe agent actions. The approach combines neural classifiers that evaluate agent behavior with symbolic rules that deterministically enforce policy decisions.

Sondera’s policy and agent control platform is currently in private beta. The company has made its open source harness and SDKs available on GitHub and invites teams to participate in early access through its website.