Grok Bot inference hooks let your security server block a prompt before the model sees it

3 weeks ago

Grok Bot inference hooks send each model request, signed, to your organization's AI security server, which allows or denies it before the model runs. A beta for every Enterprise team: Shadow mode, fail-open or fail-closed modes, what blocked members see, and what the hook doesn't cover. Plus Pages, and Apply to MCP servers for the network allowlist. https://docs.x.ai/grok-bot/inference-hooks https://docs.x.ai/grok-bot/security https://docs.x.ai/grok-bot/files-and-results https://docs.x.ai/grok-bot/teams-and-enterprises https://docs.x.ai/developers/release-notes.md

Ask about this presentation

Chapters

  1. 0:00Inference hooks
  2. 0:14Gathered content
  3. 0:27Shadow mode
  4. 0:41Fail open or closed
  5. 0:54What members see
  6. 1:06Inputs, not actions
  7. 1:21Coverage
  8. 1:35Pages
  9. 1:50Apply to MCP servers
  10. 2:07Your security server
Show transcript
  1. Grok Bot can now send each model request to your organization's AI security server. The request is signed, and your server allows or denies it before the model runs. It's a beta for every Enterprise team.

    Slide text
    Grok

    Grok Bot inference hooks let your security server block a prompt before the model sees it

    1 Grok Bot model request
    2 Your security server allows or denies
    3 The model runs
    docs.x.ai/grok-bot/inference-hooks · Beta, all Enterprise teams
  2. The check covers what the Bot picks up, as well as what members type. Web pages, app results and command output reach your server before the next model call, so your data loss policy can stop them.

    Slide text
    Grok

    Web pages and app results get checked too

    Prompts Web pages App results Command output
    docs.x.ai/grok-bot/security
  3. A new configuration starts in Shadow mode. Your server sees every covered request and denials go to the audit log, while members keep working. Their conversations still reach your server, so tell members first.

    Slide text
    Grok

    A new hook starts in Shadow mode and blocks nothing

    Shadow mode
    Your server Gets every covered request
    A deny Goes to the audit log
    Members Keep working
    docs.x.ai/grok-bot/inference-hooks
  4. Once the verdicts look right, you pick a mode. Allow the Request lets traffic through when your server times out or errors, and Block the Request stops it. A failure never counts as a deny.

    Slide text
    Grok

    You choose what happens when your server can’t answer

    Mode On a deny When the check fails
    Shadow Lets it through Lets it through
    Allow the Request Blocks Lets it through
    Block the Request Blocks Blocks
    docs.x.ai/grok-bot/inference-hooks
  5. On a deny, the sender sees Blocked by your organization's policy, with your server's reason. Edit puts the text back so they can send something different, and Reset clears the conversation.

    Slide text
    Grok

    A blocked member sees your reason and can edit the message

    Blocked by your organization’s policy
    Reason Written by your server
    Edit Puts the text back in the message box
    Reset Clears the conversation
    docs.x.ai/grok-bot/inference-hooks
  6. A tool call reaches your server only after it has run, and that includes the Bot's replies to members. So a deny stops the next model call. Keep the network policy and Auto Review in place to control actions.

    Slide text
    Grok

    A deny stops the next model call, after the tool has run

    1 The Bot runs a tool
    2 Its result joins the next request
    3 Your server judges it
    docs.x.ai/grok-bot/inference-hooks
  7. Coverage stops at Grok Bot. Cloud Agents, editor agents, the CLI and the Agent SDK go to the model unchecked. And with a hook on, Grok Bot skips some early model calls, so replies can start more slowly.

    Slide text
    Grok

    Hooks cover Grok Bot only, and replies can start more slowly

    Covered
    Grok Bot conversations
    Subagents and routines
    Outside the hook
    Cloud Agents
    Editor agents, CLI, Agent SDK
    docs.x.ai/grok-bot/inference-hooks
  8. A Page is a styled report, dashboard or tracker the Bot builds, like signups for the last 30 days. Self-serve plans can share one publicly, Enterprise plans with chosen people and teams, and routines can keep it updated.

    Slide text
    Grok

    Your Bot can now build a Page you can share

    Reports Dashboards Trackers Charts and heat maps
    docs.x.ai/grok-bot/files-and-results
  9. Grok Bot reaches connected apps from Cursor's servers, so a network allowlist alone misses them. Turn on Apply to MCP servers and every app, GitHub and X included, must match an allowed host. It's off by default.

    Slide text
    Grok

    A network allowlist can now cover connected apps too

    Apply to MCP servers
    Covers Every connected app, GitHub and X included
    Default Off
    Plan Enterprise
    docs.x.ai/grok-bot/security
  10. Setup steps, the request format and the audit events are in the Grok Bot inference hooks guide. Which security server would your team point Grok Bot at?

    Slide text
    Grok

    Which security server would your team point Grok Bot at?

    Setup guide: Grok Bot inference hooks
    docs.x.ai/grok-bot/inference-hooks

What happened in AI today, in your inbox

One short email each evening. Read it and you’re caught up.

Built with ❤️ in New York