What happened
On Oct 21, BeyondTrust announced GA of AI Agent Security on the Pathfinder Platform (Windows/macOS endpoints), giving real-time control over individual AI agent actions without restricting the same actions for humans. It ships alongside Phantom Labs research showing ~1.28M shell commands from ~132K prompts in four weeks (~10 per prompt) and 85% of coding-agent users having at least one MCP server (avg 4.35).
Why it matters
Concrete evidence of the agent/command amplification gap and the MCP-sprawl attack surface. This is a privilege-centric incumbent shipping a production agent-security product rather than a beta, aimed at constraining agent actions at the endpoint.
Applicability
BeyondTrust/Pathfinder customers and enterprise teams governing agentic coding tools (Claude Code, Codex, MCP servers) should evaluate it for per-agent action policy and privilege control.