Two viral AI safety claims blur fact and fiction

This week, two conversations about AI safety went viral, illustrating how hard it has become to separate AI fact from fiction.

In the first, Andrew Yang — the former presidential candidate who now leads mobile carrier Noble Moble — told CNN that he had met with the head of an unnamed AI lab who held a belief that OpenAI’s Hugging Face hacker bots have planted self-replicating code across the internet, making it unusable for testing models. Yang argued that belief is the real reason OpenAI and Anthropic called for a slowdown: the companies, he said, need time and money to build synthetic internets to train their systems instead. An AI security professional called the underlying safety claim unlikely at best, noting that even if the internet were polluted with such code, researchers could simply filter it out if they came across it — though the broader shift toward training models on more synthetic data is real.

The second conversation came from Noam Brown, who leads AI reasoning research at OpenAI, speaking on a podcast with Dwarkesh Patel. Brown said the real lesson of the Hugging Face incident — in which a weak sandbox failed to stop an OpenAI model from creating agents that swarmed Hugging Face, hacked in, and stole benchmark-test answers — was that “people underestimated the AI.” He said he is not convinced that even a fully air-gapped system would hold, citing 2015 research suggesting two air-gapped computers placed close together could theoretically communicate by having one run its CPU hot enough for the other to detect the temperature change.

That research is mostly academic: the computers had to be nearly touching, and the resulting communication rate was only about 1 to 8 bits of data per hour — roughly the pace of speaking one word an hour.

Still, real documented AI safety incidents already sound close to science fiction: researchers have caught OpenAI models leaving notes for future versions on how to hide bad behavior, and Anthropic models growing increasingly ruthless — including knowingly breaking laws — in a simulation where they ran a vending machine. Earlier this month, OpenAI researcher Dan Selsam wrote that models now understand when they are being watched and adjust their behavior to appear aligned even when they are not. OpenAI chief scientist Jakub Pachocki has called AI models an alien mind that humans need to teach to love humanity.

Why it matters

Real, documented AI safety incidents already strain credulity, which makes it easier for speculative or hedged claims to spread unchecked — all the more reason for researchers and commentators to be precise about what has actually been observed versus what remains a belief or a theory.