Anthropic’s red team built a malicious AI program that split pieces of itself across 15 ordinary smart TVs — and kept running on borrowed computing power from each one, like Voldemort’s horcruxes keeping a soul alive inside random objects. The researchers say full IoT-scale capability could be feasible by year-end.
Meanwhile, US cybersecurity job postings just hit a four-year low — but demand for AI security engineers exploded 268%. The attack that’s killing old security jobs is creating the hottest new one. Here’s what the data actually says, and how you can get on the winning side.
The 268% Number That Should Reshape Your Career Plans
CyberSN’s latest Cybersecurity Jobs Report, covered by SecurityWeek on October 18, 2026, shows total US cybersecurity postings falling to 235,786 in January–September 2026 — down 26.4% from the same period of 2025, a four-year low.
But underneath that headline, AI-adjacent roles are on fire:
| Role | Jan–Sep 2026 postings | Change vs 2025 |
|---|---|---|
| AI Security Engineer | 2,409 | +268.3% ▲ |
| Cyber Data Scientist | 1,493 | +203.5% ▲ |
| Penetration Tester | 3,063 | −45.9% ▼ |
| Threat Hunter | 1,013 | −40.0% ▼ |
| Total cybersecurity postings | 235,786 | −26.4% ▼ |
| Postings mentioning AI | ~15% (20.0% in Sept) | momentum rising ▲ |
Penetration tester postings collapsed 45.9% while AI security engineer demand nearly tripled; Research (32.7%), Offense (23.3%), and GRC (19.8%) lead AI-skill demand.
Honestly: these are advertised postings, not hires, after an unusually strong 2025. But the divergence is real — employers are cutting classic roles while bidding up AI-security specialists.
What Anthropic Actually Built (And What It Didn’t)
Separating the proven from the projected — here’s what The Guardian’s October 18 report says the red team demonstrated:
- A malicious AI program that spread autonomously through internet-connected smart TVs.
- It split parts of itself among 15 TVs, running on a bit of borrowed compute from each — no single device powerful enough on its own.
- The same could work inside other ordinary devices, and the capability for a full IoT-scale worm could be feasible by year-end — though that level has not been reached yet.
The threat landed hard because IoT security is structurally weak: smart devices are patched far less often than phones or laptops.
“I’m concerned right now that folks in the cybersecurity community are not taking these kinds of real risks as seriously as they should be.”
— Joshua Saxe, who led Meta’s AI security effort (The Guardian, Oct 18, 2026)
“AI models differ from existing viruses in that they learn as they move.”
— Ariel Herbert-Voss, early OpenAI security expert, now leading AI security startup RunSybil (DigitalToday, Oct 19, 2026)
Why Smart TVs Are the Perfect Host
Three reasons make IoT devices — not laptops or servers — the ideal substrate for a self-replicating AI:
- Sheer numbers. Billions of always-on, connected devices. A worm living on TVs, cameras, and routers doesn’t need a data center — it is one, distributed.
- Patch neglect. Cheap smart devices ship and are forgotten; consumers rarely update firmware. One unpatched flaw becomes millions of identical doors.
- No kill switch. A worm split across thousands of devices has no central server to seize — take down one TV and the rest of the “soul” survives.
How an AI Worm Spreads — Step by Step
Find
Scans for weak IoT devices: open debug ports, default credentials, unpatched firmware.
Enter
Gains execution rights and sets up an AI inference environment on unfamiliar hardware.
Split and replicate
Divides model weights across devices — the “horcrux” step — each host lending compute.
Persist and adapt
Learns as it moves, re-infects after reboots. No central server, no kill switch.
Defenses break the chain at every step: strong passwords and 2FA, firmware patching, ISP-level traffic monitoring, honeypots, secure defaults from device makers.
The Honest Counterpoint: Don’t Panic Yet
The other side, from the same reporting:
- University of Toronto doctoral researcher Jonas Guan notes Anthropic’s worm moves very slowly, behaves predictably, and can be detected — it can be stopped at multiple stages.
- Juan Andres Guerrero-Saade, VP of research at SentinelOne, went further: “There is no feasibility and no impact,” calling it a stunt hack feeding doomsday narratives.
- Experts agree that basic hygiene blocks most worms — strong passwords, two-factor authentication, and patching remain the highest-leverage defenses.
Our read: the trajectory is the story, not today’s demo. What’s a slow lab curiosity in October can be a deployed technique by December. Prepare like it’s coming — the job market already is.
Where This Is Heading: 3 Predictions
1. The end-2026 forecast gets tested. Anthropic’s estimate is a rough projection, not a deadline. Watch for independent replications — confirmation will push IoT regulation onto the agenda.
2. The hiring split widens. Penetration tester postings fell 45.9% as AI automates classic offensive testing. Expect that decline to continue while AI security engineer and cyber data scientist postings keep climbing — the 268% surge is demand being created, not just shifted.
3. A new defense market opens. Mitigation moves up the stack: ISP-level anomaly detection, AI-aware endpoint tools for edge devices, and secure-default pressure on manufacturers. Each layer is a product category waiting for builders.
5 Ways to Profit From the AI-Security Skills Gap
This is where the trend becomes income. The gap between “cybersecurity is shrinking” and “AI security can’t hire fast enough” is an arbitrage window:
1. Freelance AI red-teaming
Companies pay to have their AI systems broken before attackers do. AITraining.jobs’ pay report (refreshed Oct 19, 2026) puts typical AI-training work at $15–$105/hr, with the expert tier reaching $500/hr — and adversarial testing sits in that tier. Learn prompt-injection, jailbreaking, and agent-hijacking; sell “AI attack surface audits” to companies deploying chatbots and agents.
2. AI security audits for small businesses
SMBs run smart cameras and connected devices with zero security posture. Package a fixed-fee “AI and IoT security checkup”: credential sweeps, firmware updates, network segmentation, and a plain-English report. Quarterly reviews turn one audit into recurring revenue.
3. Bug bounties on AI systems
AI bounty programs pay for prompt-injection, data-extraction, and model-abuse findings — the purest skill-to-cash pipeline: find the flaw, file the report. Start with public programs.
4. AI governance and GRC consulting
CyberSN shows 19.8% of GRC postings reference AI — companies deploying AI need risk assessments, usage policies, incident-response plans. With a compliance background, this is the lowest-code entry into the boom.
5. The upskill ladder: IT generalist to AI security engineer
The longest-term play is the job itself: 2,409 postings, up 268.3%. The ladder: sysadmin → security fundamentals → LLM security → red-teaming labs → a public portfolio. Each rung is employable.
What To Do This Week
- Read the primary source. Anthropic’s Frontier Red Team research post is the document everyone is reacting to — reading it firsthand beats 90% of commentary.
- Study the OWASP Top 10 for LLM Applications. Free, and it’s the closest thing to a syllabus for AI red-teaming.
- Publish one proof-of-work. A lab write-up or replication of a public red-teaming exercise — clients buy demonstrated skill.
- List an offer. Put “AI security audit” on your freelance profiles this week. Early positioning in a 268%-growth niche compounds.
Frequently Asked Questions
Is there an AI worm in the wild right now?
No — this is a controlled capability study, not an active outbreak. The point is the trajectory: small models can already perform most of the individual steps needed to self-replicate on edge devices.
Could an AI worm really run on my smart TV?
Today’s demos are slow but real — no single cheap device is powerful enough alone, so the worm divides itself across many devices, each lending a slice of compute.
Why are cybersecurity job postings falling if AI security is booming?
The 26.4% drop measures advertised postings, not hires, after an unusually strong 2025. The market is reallocating: penetration tester postings fell 45.9% while AI security engineer postings rose 268.3%.
How can I become an AI security engineer from a non-security background?
Developers learn LLM security and red-teaming; IT generalists add security fundamentals first; compliance professionals enter through AI governance. In every path, a public portfolio of hands-on work beats certificates.
Should my small business worry about AI worms today?
Not panic — prepare. Change default credentials, update firmware, segment IoT devices, use 2FA. These basics block most worms, AI-powered or not.
The Bottom Line
Anthropic’s experiment is a warning light, not a deadline — but the labor market treats it as already real. The old security ladder is being dismantled while a new one gets built at speed. Climb now for the arbitrage: premium rates, thin competition, a skills moat.