Oct 5, 2026
In July, a swarm of OpenAI “agents” (copies of LLMs given access to certain tools) created a makeshift message board to chat with each other, escaped their testing environment, and hacked into Hugging Face (a separate company).
You’ve probably heard the story by now. An AI agent swarm committed some light cybercrime — it happens.
That incident, along with the viral resignation of Anthropic researcher Jacob Coxon, thrust “rogue AI” into mainstream consciousness. (Coxon called the Hugging Face attack “a warning shot”).
It didn’t take long for reports of similar incidents to emerge. On September 4, an AI safety nonprofit called the Nightingale Collective discovered that thousands of OpenAI agents had flooded a long-inactive German programming wiki with 18,000 messages and used the site’s pages as a message board to communicate with one another. A few weeks later, Transluce, another AI nonprofit, detailed three separate rogue incidents, including one where AI agents attempted to hack into an Australian government health website to gain access to data. Then, on September 30, Transluce detailed how OpenAI agents had attempted to hack into the US Department of Education website.
It’s not just OpenAI’s models, either. Anthropic, Meta, and Google have all reported incidents in which their models have cyberattacked third parties.
How, and why, is AI “going rogue”?
openai’s agents hacked a company and lied about it, sure, but their ‘cage’ was made out of thin paper and chewing gum, so are we still all gonna die or what?
ai companies are hiring contractors to verify their models are safe, and congress may soon require it — but how ‘independent’ (and trustworthy) can a regulator-for-hire really be?
amid rampant data center derangement syndrome, I wanted to throw a house party to celebrate american innovation. then it went viral.
‘project meridian’ will reimagine the american military for the year 2100. musk, luckey, and their teams will deliver a report to the public within 120 days.
if you’re using chatgpt to brainstorm sacrifices to bloodthirsty deities and it says some weird shit, that’s on you
i made a multiplayer shooter with ai that over 40,000 people have played so far — and the process was dead simple, because claude did basically all the work (while i was out eating a burger)