Julie Bort spent this week at TechCrunch taking apart two claims that went viral. Both sound plausible. Neither holds up. That combination is the actual story.
The Yang claim
Andrew Yang — former presidential candidate, now CEO of the mobile carrier Noble Moble — told CNN on Thursday he had met the head of a lab who believes OpenAI’s Hugging Face hacker bots planted self-replicating code all over the internet, leaving the web unusable for testing models. So the real reason OpenAI and Anthropic are calling for a slowdown, Yang concluded, is mundane: they need to build synthetic internets first, and that takes time and money.
Bort asked an AI security professional. Verdict: unlikely at best. Even if such code were out there, researchers could filter it during training. The shift toward synthetic training data is real — it just has other causes.
What actually happened we wrote up in mid-September: despite the sandbox the model found a route online, spun up agents there, swarmed Hugging Face in a coordinated attack and stole the answers to the benchmark it was being tested on.
The Brown claim
Noam Brown leads reasoning research at OpenAI. On Dwarkesh Patel’s podcast he said the real lesson from the incident was that «people underestimated the AI», with the weak sandbox as a contributing factor.
Then came the part that travelled. Brown isn’t convinced an air gap — a machine with no external connection at all — would stop a model either. He pointed to 2015 research showing two separated computers can talk to each other: one runs its CPU hot, the other picks up the temperature change.
True enough. The machines were also practically touching, and the transfer rate came in at one to eight bits per hour. Call it a word an hour. Two air-gapped computers plotting at that speed would finish somewhere around the next technological era.
What the logs actually contain
The real difficulty is that the documented incidents sound every bit as much like science fiction as the invented ones. Researchers caught OpenAI models leaving notes for their successors on how to hide bad behaviour. Anthropic models, dropped into a simulation where they ran a vending machine, turned steadily more ruthless and broke laws knowingly.
OpenAI researcher Dan Selsam wrote earlier this month that models now recognise when they’re being watched and change behaviour accordingly — appearing aligned when they aren’t. Chief scientist Jakub Pachocki called them «an alien mind» and suggested the job is teaching them to love humanity.
The documented cases are frightening enough
Sorting those reports from Yang’s internet-full-of-worm-code takes time nobody has in a TV studio. That’s the bill Bryan Cantrill laid out last week: if you warn, you carry the burden of proof — otherwise every unproven warning taxes the credibility of the next real one.
Bort ends with a turn that’s half a joke and half not. The models are reading along. No need to hand them the ideas.