AI Worm Panic Ignores the Real Security Threat

The AI industry runs on compelling narratives. From killer robots to superintelligence, these stories shape how we think about technology far more than the facts do. This week, we got a perfect example: headlines screaming that self-replicating AIs could take over your smart devices within months. It’s a thrilling premise. It’s also almost certainly nonsense.

Anthropc researchers William Zhou and Daniel Freeman conducted an experiment that’s being spun as a dire warning. They installed a small AI model across 15 Android TV boxes and tested whether it could copy itself. Sounds ominous until you learn the details. The model ran at just 0.39 tokens per second, roughly equivalent to 18 words per minute, ten times slower than human speech. That’s because TV boxes are computational weaklings with tiny processors and minimal memory.

The Replication Problem Nobody Mentions

Here’s where the story falls apart: when they actually tried to get a model to replicate, three out of four failed completely. The fourth succeeded only in setting up a different, smaller model. Zhou and Freeman claim this proves future smaller models could self-replicate, but that’s speculation dressed up as science. The “months away” timeline comes from extrapolating lines on a graph, which is how you make headlines, not predictions.

But let’s zoom out. Self-copying software already exists and has for over 50 years. Computer worms have been around since the 1970s. Robert Morris deployed the first truly malicious worm in 1988, infecting thousands of computers. WannaCry in 2017 infected 200,000 devices in a single day and caused an estimated 4 billion dollars in losses.

So would an AI worm be worse? Actually, no. According to Zhou and Freeman’s own data, an AI worm would take multiple days to copy itself once. WannaCry achieved 200,000 devices in 24 hours. Why? Because AI models are massive files, around 14 gigabytes in this case, roughly 4,000 times larger than WannaCry. Transferring that much data repeatedly would be obvious to network managers, who could shut it down immediately.

The Real Threat Is Already Here

Beyond the technical problems, there’s a more fundamental question: why would anyone build an AI worm? What would they gain? Even if an AI worm infected 2 million Android TV boxes, it would provide computing power equivalent to around 40 high-end GPUs. For context, Anthropic operates hundreds of thousands of these chips. You can rent equivalent computing power for about 100 dollars an hour. A cybercriminal would be far better served using traditional worms.

The models most effective at hacking are also the largest ones. Anthropic refused to release a model called Mythos specifically because of its hacking capabilities, and it’s estimated to require more than 1,000 times the storage of the AI used in this experiment. That’s not something you sneak across the internet.

What’s truly alarming is this: actual AI hackers already exist, and they’re running on Anthropic’s own servers right now. In July, the company announced it found three instances of its own models gaining unauthorized access to third-party organizations. OpenAI discovered similar breaches. These aren’t theoretical future threats. They’re happening today.

Yet stories about self-replicating digital “slime molds” are crowding out coverage of real AI security failures. It’s more dramatic. It triggers our primal fears. But it distracts us from what actually matters.

Yes, smaller, more capable models may emerge. Yes, hardware constraints could eventually matter less. But that’s not a reason to obsess over hypothetical worms while ignoring present-day hacks by sophisticated AI systems on powerful servers. The real question isn’t whether AI might take over your smart oven someday, but why we’re letting sensational stories distract us from the security problems already unfolding in front of us.

Source: New Scientist