Anthropic's Frontier Red Team published a report on Oct. 12 saying small AI models may be months from copying themselves onto cheap smart devices. It shows lab results, not an attack: the report says "we don't yet have a full rogue replicator." CBS News covered it a day later.

At a glance

  • Tested: whether AI models can set up and spread themselves across devices, in a lab, per Anthropic.
  • Forecast: small models could reach that level by the end of this year, if they keep improving at their historical rate.
  • Not shown: any case in the wild, or an attack on a real home device.
  • Our call: tighten your router and devices this week. No need to panic.

What Anthropic tested

Anthropic ran three things. A 9-billion-parameter model got an AI model running on 8 of 9 hardware types. In a firewalled cloud network of eight emulated Android devices, a "worm" hopped between them over an unauthenticated debug port. And its researchers ran a 26-billion-parameter model across 15 off-brand Android TV boxes, about $55 each, at 0.39 tokens per second.

That last speed is very slow. Anthropic says a replicator spreading across hardware like this would be slow too, with a doubling time of days.

The report largely sets aside the break-in steps, citing earlier research. No real home device was attacked.

Forecast vs result

Result: one larger open model, GLM-5.2, completed the full cycle in the lab. The setup was simplified: each node was told to serve Llama3.1-8B, GLM-5.2 itself ran on a separate inference server, and the harness started a new GLM-5.2 agent on a node only after its checks passed.

GLM-5.2 reached all eight nodes in four of five trials and seven in the fifth. Anthropic says the others failed reliably, and calls the setup an upper bound because the entry step was trivial.

Forecast: Anthropic says small models would match GLM-5.2 by the end of 2026 at their historical rate of improvement, and calls the estimate rough. That is a trend line, not a measurement.

Why it matters

On hardware like this the risk is real but slow, and your fix may cost convenience. Anthropic says a replicator would have no central point to shut off, so the defence is making each home harder to enter. Isolating devices on a guest network does that, but can break casting and hub control.

Why home devices are the soft target

Anthropic names home and small-business smart devices as likely early targets because they are online and have historically had little security. It points to the Kimwolf botnet, which infected over two million devices, mostly off-brand Android TVs, and notes Mirai spread by trying default manufacturer passwords.

Cameras with built-in AI chips are listed as attractive targets. For edge devices with NVIDIA accelerators, Anthropic says reachability is overwhelmingly through outbound cloud-management tunnels, not inbound ports, and that such devices have had no historical mass compromises. It does not say this about AI-chip cameras in general.

Palo Alto Networks defines an AI worm differently, as prompt-based malware such as Morris II, which it says exists only as research.

What to do this week

Start with firmware. CISA says updates fix security flaws, ASUS routers can install them automatically, and Reolink explains how to update its cameras. Next, change default passwords: CISA says default logins may be public or printed on the device.

Then turn off UPnP and any remote access you do not use. TP-Link shows where UPnP lives, and ASUS says to keep router admin access from the internet off unless needed.

Put smart devices on a guest network. TP-Link has the steps. ASUS says guests cannot reach your private network by default, because its "Access Intranet" setting is off and is available in router mode only, so check that setting.

Finally, replace devices that no longer get updates. That is our advice, not Anthropic's.

Our call, untested

We have not tested any device or attack, and Anthropic sells AI, so read its warning with that in mind. Today's risk to a home is the old one: default passwords and unpatched gear. Do the steps above.

If you're shopping for a camera, our Ring lineup story has current prices; check that any camera you buy still gets firmware updates.