Show HN: Talos – An AI agent with a permission kernel between model and shell

Talos is an AI agent that implements a permission kernel between the model and shell, controlling what commands and actions the AI can execute. The system uses a policy-based decision engine to evaluate requests before they run, logging verdicts with reasoning, and employs multiple safety layers including sandboxing, human approval requirements for irreversible actions, and a manifest-based tool system that prevents undeclared operations. All computation happens locally in the browser with no data sent externally, backed by adversarial testing against 179 cases on every update.