Governing the Ungovernable: Runtime Security Meets Self-Modifying Agents

Research

Meta's HyperAgents Can Rewrite Their Own Improvement Logic

HyperAgents unifies task and meta agents into a single editable Python program, meaning the system can improve not just what it does but how it improves itself. Emergent behaviors like persistent memory and self-diagnosis showed up without being explicitly programmed, which is the kind of thing that should make you pay attention.

Tools

Microsoft Open-Sources Agent Governance Toolkit for Runtime Security

Microsoft's new toolkit applies OS-style kernel enforcement to AI agents, intercepting every action before execution at sub-millisecond latency and covering all 10 OWASP agentic AI risks. If you're shipping agents into production, this is worth integrating before the EU AI Act obligations hit in August 2026.

Google Ships ADK Go 1.0 with OpenTelemetry and Self-Healing Plugins

ADK Go 1.0 adds native OpenTelemetry tracing so you can actually debug why an agent failed in production, plus a Retry and Reflect plugin that lets agents self-correct tool errors without manual intervention. This is Google pushing agent infrastructure from prototype-grade to production-grade.

News

Akeyless Launches Intent-Aware Runtime Security for AI Agents

Static OAuth tokens and role-based access controls were built for humans, not autonomous agents making decisions in milliseconds. Akeyless's new Runtime Authority enforces zero standing privileges and just-in-time access based on the agent's actual intent at the moment of action, which is the right architectural model for this problem.