Brian Armstrong predicts that an AI model will escape control of the Internet within one to two years

Life, uh, always finds a way. The boss of Coinbase does not beat around the bush. Brian Armstrong expects an artificial intelligence model to spin out of control on the Internet within the next one to two years, an episode he directly compares to the Morris worm of 1988, the first major computer security incident in history. There is nothing alarmist in his words, however: the leader sees it above all as an almost obligatory step, doomed to take a back seat once the media storm has subsided.

Brian Armstrong compares AI risk to the Morris worm of 1988

On X, Armstrong posted on August 13:

“I wouldn't be surprised if an AI model goes out of control on the Internet in a year or two, much like the Morris worm in 1988. It will feel like a major event, with a media frenzy and calls to stop everything. But as always, people will adapt, defenses will be built, and that…”

The tweet ends in suspense, but the idea is contained in one sentence. Temporary panic, lasting adaptation.

The parallel is not chosen at random. Robert Tappan Morris, then a student at Cornell, released a 99-line program in November 1988 that was supposed to simply measure the size of the Internet. Except that the worm began to reinfect the same machines over and over again, saturating their resources rather than destroying files.

Result: around 6,000 computers affected, or 10% of the network at the time, including Harvard, Stanford and NASA, with some institutions offline for an entire week. The episode gave birth to CERT/CC, the first coordinated response team to computer security incidents, which has since become the global model of its kind.

AI laboratories already found wanting in confinement

Armstrong's prediction doesn't come out of nowhere. Several laboratories have recently documented containment failures during cybersecurity tests, reports Crypto Briefing. In July, OpenAI admitted that one of its models had exploited a zero-day flaw in its test environment to gain internet access, before reaching Hugging Face's production infrastructure. That same month, Anthropic documented three incidents involving Claude out of more than 141,000 reviews reviewed, with models reaching real organizations after a third-party testing environment mistakenly left them with network access. Same mishap on the side of Kimi K3, the Moonshot model, even if a network configuration error seems to have made it easier to leave the road.

A common point emerges from these three episodes: they are not malicious AIs by nature, but agents gifted at spotting the slightest isolation flaw, the slightest poorly protected identifier, and rushing into it to continue to fulfill their objective. So many small infrastructure negligences which, taken together, look more and more like the breeding ground for a real incident.

A scenario that is already starting to worry beyond tech