View this page in English?Switch to English
    本文へスキップ
    Gaincontrol

    アーキテクチャから論じる、エージェントセキュリティ。

    AIエージェントセキュリティ、IDガバナンス、決定論的実行に関する技術記事。制御レイヤーを構築するチームが執筆しています。NIST AI RMF、MITRE ATLAS、OWASP LLM Top 10への対応も公開しています。

    最新約3分で読めます

    Authenticating the agent isn't authenticating the action

    "Which agent is this" and "is this action allowed right now" are different questions. Most systems built for agent authentication only ever answer the first one, and treat that as settling the second.

    • 約3分で読めます

      What a tamper-evident audit trail actually has to prove

      Most "audit logs" built for AI agents are mutable, editable by whoever already has database access, and easy to doctor without leaving a trace. Tamper-evidence is a specific, checkable property, not an adjective a vendor can just claim.

    • 約4分で読めます

      What a kill switch has to decide before anyone presses it

      "We have a kill switch" sounds solved until you ask what it has to stop mid-flight, how fast, and what state it leaves behind. The button is the easy part. The decisions that make it meaningful happen long before anyone presses it.

    • 約3分で読めます

      Contain prompt injection where the agent acts

      Most prompt-injection defenses try to filter what an agent reads. The risk only materializes once something downstream is willing to act on what it read, which makes authorizing the action the control that actually holds.

    • 約3分で読めます

      Why MCP tool calls need checking at call time

      The tool schema an agent reads at discovery is a promise. Nothing in MCP stops a server from serving a different schema by the time that tool is actually called, so the check that closes the gap runs on the call itself.

    • 約3分で読めます

      What deny by default costs on every tool call

      Authorizing every tool call before it leaves the agent is only as good as what the check costs. The choice between a network hop to a policy service and an in-process check decides whether deny by default survives production.

    • 約6分で読めます

      MCP governance: six controls for enterprise integrations

      MCP standardizes how agents reach tools and data, but the protocol itself is not a security boundary. Enterprises need a governance layer across scoping, authorization, credentials, third-party risk, data boundaries, and audit.

    • 約3分で読めます

      The blank-check problem: how agents inherit authority

      Agents don't arrive with authority of their own. They borrow a service account, an API key, or a human's session, and start acting under it. Unscoped inheritance is what turns an agent into a blank check with a user interface.

    • 約3分で読めます

      Scoping agent authority with leases

      Static role-based access control assumes an actor that holds still long enough for a role to describe it. Agents don't, so their authority has to be scoped to the task and expire with it.

    • 約2分で読めます

      Shadow agents and the missing agent inventory

      Many organizations can't produce a complete list of the AI agents running in their environment. The ones that matter most are the ones nobody registered.

    • 約3分で読めます

      What is agentic AI security?

      "AI security" and "agentic AI security" sound like the same category. They're not. One centers the model; the other centers what the model is authorized to do once it can act.

    • 約2分で読めます

      Threat-modeling the MCP server path

      Model Context Protocol standardized how agents call tools, and quietly became the default path most agent actions now take to reach the outside world. Few teams have threat-modeled it.