The Agent Security Stack is Finally Catching Up to Agent Capabilities

Research

Meta's HyperAgents Can Rewrite Their Own Improvement Logic at Runtime

This isn't just an agent that improves at tasks, it's an agent that rewrites the code responsible for how it improves. The emergent behaviors like self-diagnosis and adaptive memory that were never explicitly programmed are the part worth watching closely.

Tools

Microsoft Open-Sources the Agent Governance Toolkit with Sub-Millisecond Policy Enforcement

Seven packages covering everything from cryptographic agent identity to compliance mapping against the EU AI Act, and it slots into LangChain, AutoGen, and CrewAI without a rewrite. The p99 latency under 0.1ms for policy enforcement means you can actually ship this in production without a performance excuse.

Akeyless Ships Intent-Aware Runtime Security for AI Agents with Zero Standing Privileges

Static OAuth tokens and RBAC were built for humans and services with predictable access patterns, not agents that need to hit a dozen systems in a single autonomous run. JIT access plus real-time intent evaluation is the right mental model here, and Akeyless is one of the first to productize it at scale.