Beyond Static Sandboxing: Learned Capability Governance for Autonomous AI Agents

Collection of recent research (Apr 2026) covering AI-agent security, privacy-preserving LLM inference, cryptographic protections, benchmarks and tooling for security automation, and vulnerability discovery for blockchain bridges. Key contributions: Aethelgard — a learned, layered governance framework that enforces least-privilege for autonomous agents to reduce capability overprovisioning; practical lightweight block-cipher payload encryption for real-time CAN nodes to hinder semantic reverse-engineering of automotive signals; SIR-Bench — a 794-case benchmark (with OUAT) for evaluating true-f1