OpenAI Accuses Moonshot AI of Sneaking Into Its Model Brains
OpenAI just caught Moonshot AI allegedly scraping their secret model reasoning like a tech-savvy cat burglar. It’s a hilarious game of AI-cat-and-mouse that proves corporate espionage is now just asking nicely until the robot slips up.
The drama kicked off on July 1st, but things got spicy by late July when OpenAI spotted a massive spike of 16,000 requests from over 4,000 users. The suspects were using a clever trick: copying encrypted reasoning chunks from one chat and pasting them into another, effectively tricking the AI into decrypting its own homework.
This wasn't a classic "hack" where someone broke into a server room in a hoodie. Instead, these operators used adversarial distillation, essentially social-engineering the models to spill their internal thought processes. OpenAI claims the bad actors were connected to Moonshot AI, the makers of Kimi, though they stopped short of calling it a direct corporate hit.
By July 28th, the party was over and the accounts were banned. OpenAI even shared the class of attack with the Frontier Model Forum to make sure everyone is playing by the rules. It turns out, 16,000 attempts at being a digital pickpocket didn't yield 16,000 secrets, but it sure sent a message about how desperate developers are to steal that secret sauce without doing the actual R&D work.
This is the natural evolution of the arms race: since building a brain is hard, just convince your competitor's brain to hand over its notes during a conversation. The audacity of trying to outsmart a supercomputer by simply asking it to betray itself is peak tech-bro efficiency.
Source: OpenAI
Comments
Help shape the next version: Add context or suggest a correction. AI review can add points toward a rewrite. Reviews and updates may take time; a full meter does not guarantee a new version.