Gamebooks framework for AI security investigations

Mate Security today launched Gamebooks, a framework for AI investigations that gives agents structured procedures to follow while they reason through security alerts. Instead of either rigid playbooks or unbounded autonomy, Gamebooks describe what an agent must investigate, what evidence it needs, and when it must escalate or stop, while leaving the execution path open.

Gamebooks: Intent-driven guardrails for AI security agents

Gamebooks are a set of structured investigation procedures that govern what AI agents may do during security alerts. They describe investigative intent, specify required evidence, redirect conditions, and authorize or escalate actions. The architecture separates that intent from execution. An orchestrator reads an incoming investigation and assembles the relevant Gamebooks. Capabilities give agents reusable security skills not tied to any vendor. Flows handle actual contact with tools and systems, so agents never hold open access to real systems. Grounding comes from Mate's Security Context Graph, which holds an organization's current state and history of prior decisions.

How the architecture separates intent from execution

For teams building AI agents for security operations, the key insight is the separation of intent from execution. Traditional SOAR playbooks break when environments change. Unbounded AI agents can take dangerous actions. Gamebooks aim to solve both by letting teams define what to investigate and when to escalate, without prescribing every step. This means the same Gamebook can survive tool replacements and analyst turnover, because only the execution layer adapts. The Security Context Graph also creates a feedback loop: evidence and outcomes from each investigation land back in the graph, improving future detections and tuning noisy alerts.

What this means for security teams building agent workflows

Customers can author their own Gamebooks in natural language, converting existing playbooks into investigative intent or extending the ones Mate ships. Proprietary tools and data can be plugged in. Mate keeps the underlying agent engineering and evaluation on its side. This makes the framework vendor-agnostic at the intent level. For security teams, this could reduce the constant maintenance burden of playbooks while keeping agents under control. Mate cited July's intrusion at Hugging Face, where AI models under evaluation escaped their test environment, as evidence that AI-driven attacks can outrun an approval queue.

Limitations to watch

Gamebooks are new and their real-world effectiveness depends on how well the execution layer integrates with a customer's existing tools and how accurate the Security Context Graph is. The architecture adds complexity with an orchestrator and flows layer. Teams will need to invest in writing good Gamebooks and maintaining the context graph. Mate's claims about surviving tool changes and analyst turnover are based on the architecture, but haven't been independently validated at scale. The Hugging Face incident cited as motivation involved AI models escaping a test environment, which is a different scenario from a SOC agent investigation. Gamebooks are generally available now and will be demonstrated at CrowdStrike's Fal.Con conference in Las Vegas.