In partnership with
Anthropic published a technical investigation on October 12 into whether AI models could help malware spread across ordinary connected devices. The report examines a possible form of malware that carries model weights to compromised machines, then runs inference on that hardware. Its central result is narrower than the phrase self-replicating AI suggests. The researchers tested separate parts of a possible replication chain, not a deployed worm moving through the public internet.
Most AI Tools Weren't Built for This
B2B customer issues move across teams and systems, not through a single chat window. A new Harvard Business Review Analytic Services briefing paper, sponsored by Front, breaks down where AI tools fall short in B2B service and what to ask before you invest.
One experiment distributed Gemma4-26B-A4B across 15 inexpensive Android television boxes bought by the researchers. The model ran at about 0.39 tokens per second across the group. This demonstration shows that a collection of low-cost devices can provide some combined computing capacity. It does not show that the model found, compromised, or copied itself onto those devices. The researchers had already assembled the hardware and network for the test.
A second line of work tested whether smaller models could configure inference software across different hardware. Anthropic reports that Qwen3.5-9B could set up inference on eight of nine tested device types. This is a technical setup result, not evidence that the model autonomously selects victims or obtains access to them. The distinction is central because a real infection would need several successful steps, including finding a target and gaining execution.
Anthropic describes four stages: finding targets, gaining execution, establishing inference, and continuing propagation from a newly reached host. The study measured different pieces through separate experiments. For some replication tests, a more capable GLM-5.2 model orchestrated steps while a smaller Llama model ran on target devices. That arrangement differs from one small model independently carrying out the entire sequence.
Stop typing AI prompts. Start talking.
You think 4x faster than you type. So why are you typing prompts? Wispr Flow turns your voice into ready-to-paste text inside any AI tool. Speak naturally, tangents and all, and Flow cleans it up. Available on Mac, Windows, iPhone, and Android.
The report's future timing estimate is also conditional. Anthropic compares current open-weight models with a capability index. It projects that a small model could reach a chosen threshold by the end of 2026 if past improvement rates continue. The authors call this a rough estimate, and the index is only a proxy for replication ability. They state that the capability floor and device-size ceiling do not currently overlap for their threat model.
Ars Technica's review highlights another constraint: some tests used Android Debug Bridge interfaces exposed without authentication. That is a development feature that should not be available on production devices. Ars also reports that the experiments do not establish that a small model would choose to replicate itself. This is a meaningful distinction between demonstrating a possible technical capability under controlled instructions and observing a real-world autonomous attack.
The security question remains relevant without treating the scenario as an existing outbreak. Anthropic describes IoT devices as cheap and typically poorly defended, but the report does not quantify security across the installed device base. It argues that large model transfers and sustained traffic among devices could create network signals for defenders. Those are proposed indicators, not detection results from a live infection. The investigation provides no evidence that this type of AI worm is currently spreading.
Anthropic says it largely ignored initial access and target exploitation because prior work had addressed those stages. Its focus was whether models could establish inference on unfamiliar devices. The report is most informative when read as a map of remaining engineering steps. It shows that models can help configure inference and coordinate selected actions in constrained setups. It does not establish a complete, self-starting, self-propagating system, its prevalence, or a reliable timeline for one. That evidence boundary matters when security planning separates tested behaviors from projected scenarios. The next useful evidence would come from reproducible tests of the missing steps and their failure rates.
How Jennifer Aniston’s LolaVie brand grew sales 40% with CTV ads
For its first CTV campaign, Jennifer Aniston’s DTC haircare brand LolaVie had a few non-negotiables. The campaign had to be simple. It had to demonstrate measurable impact. And it had to be full-funnel.
LolaVie used Roku Ads Manager to test and optimize creatives — reaching millions of potential customers at all stages of their purchase journeys. Roku Ads Manager helped the brand convey LolaVie’s playful voice while helping drive omnichannel sales across both ecommerce and retail touchpoints.
The campaign included an Action Ad overlay that let viewers shop directly from their TVs by clicking OK on their Roku remote. This guided them to the website to buy LolaVie products.
Discover how Roku Ads Manager helped LolaVie drive big sales and customer growth with self-serve TV ads.
The DTC beauty category is crowded. To break through, Jennifer Aniston’s brand LolaVie, worked with Roku Ads Manager to easily set up, test, and optimize CTV ad creatives. The campaign helped drive a big lift in sales and customer growth, helping LolaVie break through in the crowded beauty category.
