Add article: Agentic security is only as good as the device data underneath it

[allenhouchins]

Related issue: NA (drafted from a news-scan routine, no tracking issue)

What this is

A draft thought-leadership article (category: articles) pegged to Microsoft's July 27, 2026 launch of MAI-Cyber-1-Flash and Project Perception.

The argument: agentic security tooling is shipping now, and an agent that triages and remediates is only as reliable as its picture of your devices. The piece lays out what that picture needs (current, cross-platform, inspectable, able to answer artifact questions before a CVE exists), shows how to inventory the agentic tooling already running on your fleet, then argues for a narrow write path and a GitOps audit trail.

New file: articles/agentic-security-needs-ground-truth.md

Formatting

Follows the house article format: sentence-case title, italic dek, Key takeaways (6 outcome-first bullets, one per body section) immediately after the dek, CTA button before the intro, two-paragraph intro, See it live block, closing italic CTA, and meta endmatter.

The ai_tools section

The strongest section is "Start with the agents you already have," built on fleetd's ai_tools table. It turns the article's thesis back on itself: before deciding whether to let a vendor's agents act in your environment, inventory the agents already acting in it, with what reach and on whose authority.

It leans on the risk_flags column to make that concrete, since flags like mcp_shell_exec, mcp_fs_write, bypass_permissions, skip_permissions_runtime, plaintext_secret, world_readable_config, injection_markers, and hidden_unicode describe what each tool has been permitted to do rather than just that it exists.

Everything in that section (queries, column names, type values, and every risk flag named) comes from schema/tables/ai_tools.yml. The section also documents two limits, so an empty result isn't misread as an absence:

• root is required for full visibility across all user home directories
• symlinked config and binary paths are deliberately not resolved

One thing to confirm: the draft doesn't name a version number for ai_tools, since orbit/changes/47619-ai-tools-table is still pending rather than released. Please add the concrete fleetd version at publish time, or confirm you'd rather leave it unversioned.

Sourcing and honesty notes

Two other things I want a second pair of eyes on:

1. Microsoft's benchmark figures are vendor claims. The 96% CyberGym score, the half-the-cost comparison, and the "100 trillion daily security signals" figure come from Microsoft's own announcement via press coverage, not from primary documentation. The draft attributes them explicitly ("Microsoft reports…") and adds a line telling readers to weigh them accordingly. Worth confirming the numbers and the August 3 public-preview date before publish, since launch details shift.
2. How we reference fleet-mcp. The draft describes it as an open-source MCP server built by Dhruv Majumdar, Fleet's VP of Security Solutions, and says plainly that it's a community project rather than a supported Fleet product. That mirrors how the published MCP manifesto article frames it, but please confirm that's still the framing marketing wants.

Per the competitor-framing rules, Microsoft's launch is described factually with no editorializing, and the piece explicitly credits the approach ("this is not a knock on the approach") rather than positioning against it.

To avoid overlapping the published shadow AI article, this piece stops at inventory and cross-references that one for the fuller find-and-govern tour.

Before merge

• Add the fleetd version for ai_tools once the release lands
• publishedOn is set to 2026-07-30; update to the real publish date
• No articleImageUrl set (optional meta, present on about half of existing articles). Add one if you want a card image.
• Confirm the CTA destination /visibility-and-reporting is the intended landing page. I picked it because /observability and /orchestration both 301-redirect there.

Checklist for submitter

Content-only change, so most of the standard checklist doesn't apply. Trimmed to what's relevant:

• No changes file needed (website content, not a user-visible product change; the ai_tools changes file already exists at orbit/changes/47619-ai-tools-table)
• QA'd all new/changed functionality manually (n/a for prose; reviewer should read for voice and accuracy)

https://claude.ai/code/session_01RtQYtRJZGBArkR8JrCEesR

[Copilot]

Copilot wasn't able to review any files in this pull request.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Copilot wasn't able to review any files in this pull request.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review is ineligible. To be eligible to request a review, you need a paid Copilot license, or your organization must enable Copilot code review.