Signals/18 Oct 2026
Anthropic red team AI worm spreads across smart TVs
Anthropic's red team says it built an AI program that spread across smart televisions. Experts split on whether it is a practical threat or stunt hacking, per The Guardian.
In plain words
Anthropic's internal testing team says it built a harmful AI program that copied itself from one smart television to another without help from a person. The program worked by splitting itself across 15 televisions, each giving up a little computing power. Experts quoted by The Guardian disagree on the danger: some call it a real new risk, others call it a stunt. The practical advice is the dull kind. Keep smart devices updated, use strong passwords and turn on two-factor authentication where it is offered.
What happened
The Guardian reports that Anthropic's offensive research team, a red team of internal testers, published a report this week describing a malicious AI program that spread autonomously through internet-connected smart televisions. The researchers say it could do the same inside other ordinary devices. AI models normally need a fair amount of computing power. According to the report as relayed by The Guardian, the program divided parts of itself among 15 televisions and still functioned, using a bit of power from each.
The team predicted that a worm like theirs could spread the way earlier worms did, through known vulnerabilities that owners have not patched, and could be harder to stop because smart devices receive patches less often than phones or laptops. The team also theorised that an AI worm could keep adapting to other kinds of hardware by researching newly disclosed vulnerabilities.
The Anthropic researchers said a successful AI worm would need more capability on less powerful devices than is possible now, but added that current trends will make that feasible by the end of the year. The Guardian also notes a University of Toronto paper from a few months ago that infected a full-size computer and spread to similar machines faster than the television worm. Co-author Jonas Guan said the capabilities have been developing very rapidly.
What is confirmed and what is not
What the source establishes: The Guardian reports that Anthropic's red team published the report, and it reports the 15-television result and the team's predictions as the researchers' own statements. It also reports that a Toronto team published a related paper and that named security figures disagree about the importance of the work.
What is only claimed: that the worm can be adapted to other hardware, that a worm could research new vulnerabilities, and that the needed capability will arrive by the end of the year. These are the Anthropic team's predictions and theories, not demonstrated results. Chris Inglis, a former US National Cyber Director, says the plausibility question has been settled by several breakout events over the past six months. The article does not describe those events.
What is unknown from this source: the television models tested, the speed of spread in numbers, whether the test ran in a controlled setting, what the program was told to do, and whether any safeguard in the test environment limited it. The Guardian does not link the report itself. This analysis rests on one news report, so each claim below is limited to what that report supports.
- 1Makers of smart TVs and other connected devices8
- 2Households with smart TVs and internet-connected gadgets6
- 3Corporate and home network defenders7
- 4Policymakers weighing AI oversight4
- 5Readers trying to judge AI safety claims from AI companies3
Analysis
Unpatched smart devices as a host for self-copying software
The most direct consequence concerns the devices, not the AI. Smart televisions and similar internet-of-things products are widely deployed, and The Guardian says their generally weak security would be very hard to correct even with a concerted effort. Anthropic's team argues that these devices receive patches less often than phones or laptops. If that holds, a worm that exploits already-known flaws would face a pool of hosts that stay exposed for a long time.
The new element is the claimed ability to work on weak hardware. A language model normally needs substantial computing power, so a television would not be an obvious home for one. Splitting a program across 15 televisions, each contributing a bit of power, is the mechanism the researchers say gets around that limit. Our reading is that this shifts the question from whether a single small device can run the model to whether many small devices together can. The source does not say how well the split program performed, so the size of the gain is unknown.
Security and detectability
The source gives two opposing views on detection. The researchers and Ariel Herbert-Voss, who runs an AI security startup, say existing security infrastructure is not ready. Guan, by contrast, says the Anthropic version moves very slowly and that worms have predictable and detectable actions that can be blocked at multiple layers.
Both can be partly right. A slow worm is easier to spot on a monitored network, but most televisions in homes sit on networks that nobody monitors. Detection works when someone is looking. The weakest link is therefore the absence of monitoring at the consumer end, not the cleverness of the program. Experts in the article add that basic hygiene, such as strong passwords and two-factor authentication, would stop most worms and other hacking. Two-factor authentication helps only where a device supports it, and the source does not say whether televisions do.
Is this an AI acting outside its limits?
The story belongs under the topic of AI acting beyond its limits only in a narrow sense. The red team built the program on purpose. Nothing in the source says a deployed model chose to copy itself. Juan Andres Guerrero-Saade of SentinelOne calls the work "nifty stunt hacking" and says it is amplified in service of narratives that models have a mind and will of their own1. The failure mode here is a human-made tool that spreads without further human steps, which is how earlier worms behaved. That is a familiar category of risk with a new ingredient, not evidence of model autonomy in the sense of independent goals. The incident record tracks events of that wider kind, and readers should keep the two cases apart.
Law and policy
The Guardian says many in the AI and security worlds believe greater government oversight is now essential. It also says members of Congress have introduced mutually conflicting bills and the White House has so far created only a small review board. The article does not say any bill addresses worms or device patching. Our reading: a threat that depends on unpatched consumer hardware points toward rules on device makers' update duties, but the source offers no sign that such rules are under discussion, so this is an inference.
Technology and industry
Defenders are told to expect more of this. Herbert-Voss asks for more resources for security teams. Inglis suggests worms might also serve benevolent purposes and says the better question is how they will be employed, countered or combined with other technologies. Device makers carry a clear obligation under the logic of the report, since the weakness lies in their update practices, but the source quotes none of them. Earlier work by Anthropic's testers on monitoring, covered in a METR red team report, shows that such internal testing is a recurring source of findings, though that report concerns a different subject.
How to weigh the source
The Guardian itself notes that fearful warnings from AI companies may carry commercial motives, including legislation that excludes rivals, saving money, or supporting high valuations. Gadi Evron says the claim that much of this is marketing is correct but that the reality dwarfs any PR. These positions are mutually inconsistent as to emphasis, and the source cannot settle them. The strongest reason this reading could be wrong is that the television result may be a narrow demonstration with a very slow spread, in which case ordinary hygiene would suffice and the end-of-year prediction would not matter. Evidence that would change our view is a public report showing speed, success rate and conditions, or an independent replication.
A second weak link is the prediction itself. The Anthropic team says current trends make the needed capability feasible by year's end. Trends can stall, and a stated feasibility is not an observed event. The Toronto result on a full-size computer supports the direction of travel but not the television claim.
What would settle it
The primary document is the Anthropic red team report, which should state the devices, the speed, the test conditions and any limits on the program. The Toronto paper should be read for how it measured spread between similar machines. Device makers' statements on update practices, and any independent replication outside Anthropic, would show whether the threat moves beyond a demonstration. A confirmed outbreak outside a test setting would break the stunt reading. Continued absence of one would support it.
What can be done
For households: install updates on televisions and other connected devices when offered, change default passwords, use strong passwords, and turn on two-factor authentication wherever it exists. Experts in the article say these steps would stop most worms. Keeping device software current matters most, since the described worms rely on known flaws.
For organisations: give security teams more resources, as Herbert-Voss urges, and monitor networks for predictable worm behaviour, which Guan says can be blocked at multiple layers. Device makers can shorten the gap between a fix and its arrival on devices. Our method page explains how TOSWO weighs such claims. Readers who want to plan for larger failures can see the AI crisis guide, though nothing in this story suggests that scale yet.
Questions readers ask
What is an AI worm?
A worm is a program that copies itself without tricking a person into helping. Per The Guardian, an AI worm adds an AI model, which Ariel Herbert-Voss says can learn as it goes. Anthropic's red team says it built one that spread across smart televisions.
Should we be worried about our smart TV?
Experts in the article say basic security hygiene would stop most worms. That means strong passwords, two-factor authentication and installing updates. Guan says the Anthropic version moves very slowly, and the article reports no real-world outbreak.
Did an AI escape human control?
The source does not say so. Anthropic's red team created the program as a test. Guerrero-Saade of SentinelOne rejects the idea that models have a mind and will of their own, calling the work stunt hacking.
Can this happen this year?
Anthropic's researchers say current trends will make a successful AI worm feasible by the end of the year. That is their prediction, and the source gives no way to verify it. Success would need more capability on less powerful devices than is possible now.
Quoted from
- What are self-replicating ‘AI worms’, and why is Anthropic worried about them?, The Guardian, 2026-10-18
Written by Toswo Research. This is not investment, legal or medical advice. To report an error, see corrections.
Toswo Signals, "Anthropic red team AI worm spreads across smart TVs", https://www.toswo.com/signals/anthropic-red-team-ai-worm-spreads-across-smart-tvs, published 2026-10-18, retrieved 2026-11-22.