Most of the AI-and-jobs conversation right now is about the visible layer: who writes the code, who prompts the model, who gets cut when a chatbot can do the first draft. Buried in this week's news is a story about a layer nobody in that conversation is looking at — and it's worth your attention precisely because almost no one is competing for the skills it rewards.

A bug that's been sitting there since 2013

Security researchers this week detailed how thousands of internet-connected servers from major manufacturers can be remotely backdoored through their baseboard management controllers, or BMCs — the small, separate computers embedded on virtually every enterprise server motherboard, running their own OS, network stack, and IP address, independent of whatever is happening on the main machine (Ars Technica). BMCs are what let IT staff reboot, reimage, or monitor a server remotely even when it's powered off or completely unresponsive — genuinely useful "lights out" management. The catch, per the researchers, is that the underlying protocol has been flagged as a soft target since at least 2013, and some of the vulnerabilities now being exploited are more than a decade old. Ars describes it as a "pervasive, under-monitored, under-patched parallel attack surface" — not a new hole, but an old one that's been quietly ignored because it sits below where most software and security attention lives.

Agentic coding is about to make more of this, faster

The same week, Meta launched Muse Code, a terminal-based coding agent that joins Claude Code, Codex, and a fast-growing field of harnesses that write and ship software with minimal human involvement (TechCrunch). Zuckerberg's own description of it is telling: on a big enough job, it fans out into multiple sub-agents working in parallel, in isolated worktrees, building several features simultaneously. That's a real productivity gain. It's also, structurally, more infrastructure getting stood up, configured, and deployed per unit of human attention than at any point before — more cloud instances, more test environments, more servers provisioned quickly by people who may never touch a BMC setting or read a firmware changelog. Agentic tooling is very good at writing the application. It has no opinion on whether the out-of-band management chip underneath that application is running six-year-old firmware.

None of this means agentic coding tools are insecure by design — that's a separate question from the one in the Ars Technica piece. The point is narrower: the volume of infrastructure being created is scaling faster than the unglamorous discipline of patching the physical layer underneath it, and that gap doesn't close itself just because the code on top got smarter.

Why this is a career signal, not just a security story

Layer this against the layoffs still working through the industry — Monday.com's roughly 20% workforce reduction this week, tied explicitly to an AI-driven restructuring, joins a growing list of companies citing AI in job cuts (TechCrunch). It's worth being honest about what we don't know here: none of these filings break out layoffs by function, so there's no clean data saying "security and infrastructure roles are spared." That would be overclaiming. What can be said with more confidence is structural, not statistical: the roles being automated fastest are the ones where the work product is text, code, or a customer interaction an LLM can approximate. Physical and firmware-level security — knowing what a BMC is, how IPMI and its successors work, how to audit out-of-band management access, how to push a firmware patch across a fleet without bricking half of it — sits closer to the machine than to the model. It's not obviously a task current AI agents are reaching for, and it doesn't get easier to automate just because the software layer above it does.

What to actually do with this

You don't need to become a firmware engineer to act on this. A few checkable moves:

  • Ask, don't assume. If you work in IT, platform, or security at any company running its own servers (not pure cloud), ask whether BMC firmware is on a patch cycle at all. Many organizations genuinely don't know the answer.
  • Learn the vocabulary, not just the headline. BMC, IPMI, Redfish, out-of-band management — these are searchable, well-documented concepts, and understanding them puts you ahead of most people whose AI-era skill-building has been entirely prompt- and agent-facing.
  • Watch what agentic coding tools don't cover. As you or your team adopt tools like Muse Code or Claude Code for infrastructure-adjacent work, explicitly ask what's out of scope — deployment targets, firmware, physical access controls — and treat that gap as a place to build expertise rather than an oversight to ignore.
  • Don't over-rotate. This is one plausible niche among several, not a guaranteed safe harbor. Treat it as a hedge worth investigating, not a career pivot to make on the strength of one Ars Technica story.

The broader lesson generalizes past BMCs specifically: as AI gets better at the software people can see, the harder-to-automate value increasingly sits in the physical and operational layers underneath it — the parts that don't show up in a demo. That's a narrower bet than "learn to prompt," but it's one fewer people are currently making.