Funny week for the industry. On the same day Microsoft published a blog about turning “threat intelligence into decisive action” with an AI-assisted service, and the Trump administration announced an AI-powered clearinghouse for vulnerability triage, researchers quietly pointed out that a stack of revoked UEFI bootloaders had been sitting trusted on machines for years. Nobody’s cybersecurity AI caught that one. Nobody’s AI was looking.

That’s not a coincidence, it’s a pattern. The industry keeps selling smarter alert triage while the boring, unglamorous stuff, the things that don’t need a large language model to find, just sits there unpatched. If you run a security program, this week is a decent reminder that the flashiest tooling and the most basic hygiene gaps live in completely different parts of your budget, and only one of them is getting attention.

The AI Pitch Is Everywhere This Week, Again

Microsoft’s Defender Experts post leans hard on the idea that security teams have more visibility than ever but somehow feel more uncertain, because signal floods in from endpoints, identities, cloud workloads, and a dozen third-party tools nobody fully trusts. CrowdStrike is out with its own framing, something it’s calling AIDR, positioned as the next era of the industry. And the federal government just unveiled Gold Eagle, a program that lets industry and critical infrastructure operators use AI to detect, prioritize, and patch vulnerabilities faster.

None of this is inherently bad. Alert fatigue is real, and if a model can chew through ten thousand low-fidelity detections and hand an analyst the three that matter, that’s a legitimate win for threat detection and incident response workflows. But there’s a quiet assumption baked into all of this messaging: that the bottleneck is analysis. That if you just had better prioritization, better triage, better decisive action on top of the data you already have, the risk would go down.

Sometimes the bottleneck isn’t analysis. Sometimes it’s that nobody looked at the asset in the first place.

Meanwhile, A Decade Of Bootloaders Nobody Revoked

Dark Reading’s report on forgotten UEFI shim bootloaders is the perfect counterpoint. Nearly a dozen vulnerable bootloaders, already flagged, already revoked, kept working as a trusted path around Secure Boot for years. Not because the vulnerability was hard to find. Because revocation lists are tedious to maintain, firmware updates are painful to push at scale, and “boot integrity” isn’t the kind of line item that gets a press release.

This is the second time in two weeks a firmware trust-chain story has surfaced, and it keeps surfacing for the same reason: nobody’s AI clearinghouse is going to flag “this bootloader should have been blocked in 2023 and wasn’t.” That’s an inventory problem, a configuration management problem, a “does anyone actually own this” problem. No amount of threat-protection tooling fixes ownership gaps.

The Browser Extension Angle Nobody’s Automating Either

Then there’s ClaudeBleed, the flaw in Claude for Chrome that still lets a malicious browser extension abuse the assistant’s permissions to reach into your Gmail. It’s not a novel attack pattern. Browser extensions have been a soft underbelly for identity and data access for years, and bolting an AI assistant with broad page permissions onto that same browser just gives attackers a new lever to pull on an old problem.

What’s notable is that this isn’t a detection failure either. It’s a permission-scoping failure. The fix isn’t a smarter model watching for anomalous behavior after the fact, it’s not granting an assistant Gmail-reachable permissions in the first place, or at minimum isolating what extensions can touch. Defense in depth means assuming any given layer, including your shiny new AI copilot, will eventually get walked around, and building the next layer so that walk-around doesn’t matter.

So What Do You Actually Do With This

None of this means throw out your detection stack. It means stop treating “buy the AI thing” as a substitute for the unglamorous asset and access hygiene that AI tooling generally isn’t built to catch. A few things worth doing this week, none of which require a vendor demo:

• Pull your organization’s UEFI revocation (dbx) status across your fleet. If you don’t know whether it’s current, assume it isn’t, and schedule the firmware push.
• Audit browser extension permissions on any machine running an AI browser assistant or coding agent. If an extension can read page content and the assistant can act on page content, that’s a chain worth mapping.
• Treat your perimeter appliances, VPN gateways, remote access tools, the stuff facing the internet by design, as your highest-value patch targets regardless of what your AI triage tool ranks them. SonicWall’s SMA line has been exploited repeatedly for exactly this reason: it’s the front door, and front doors get hit first.
• Keep basic brute-force and abuse protection in place on anything internet-facing. A tool like IPBan Pro or an equivalent automated banning layer won’t stop a chained zero-day, but it quietly closes off the noisy, high-volume guessing attacks that would otherwise bury your logs and slow down real incident response.
• Rotate ownership reviews for firmware, boot integrity, and certificate lifecycles onto an actual calendar. “Nobody owns this” is how a revoked bootloader stays trusted for three years.

This is security hardening in the least exciting sense of the phrase: nobody writes a press release about dbx list maintenance. But it’s the layer that was failing this week, not the analysis layer the industry keeps pitching.

Sources

• Turning threat intelligence into decisive action with Defender Experts
• Trump administration unveils AI-supported clearinghouse for cyber vulnerabilities
• AIDR: How CrowdStrike Is Defining the Next Era of Cybersecurity
• Forgotten Bootloaders Expose Secure Boot Blind Spot
• Claude for Chrome flaw could let rogue extensions access your Gmail
• CVE-2026-15409, CVE-2026-15410: SonicWall SMA 1000 zero-day vulnerabilities exploited in the wild