OpenCode breach leaks developers' OpenAI, Claude and Google API keys

Aakash Gupta @aakashguptaOriginally posted 2026.10.18

Open original on X

A bunch of developers woke up this morning to AI charges they never made. Somebody broke into OpenCode, a popular AI coding app, and grabbed the keys people use to pay for OpenAI, Claude and Google. The company says the whole thing was part of a large scale operation to copy the AI models themselves. The timing is what got me. Five days earlier, Anthropic put out a post saying it had cut the amount of data these copycats could pull out of Claude by 90 to 99%. Here's how the copying works, because everything else hinges on it. You hit a top model like Claude or GPT with millions of questions, save every answer and especially the step-by-step reasoning, and train your own cheaper model to imitate it. The industry calls it distillation. Back in February, Anthropic said DeepSeek, Moonshot and MiniMax had done this more than 16 million times through about 24,000 fake accounts, and one proxy network was running over 20,000 of those accounts at once. So the labs started going after the fake accounts. Anthropic built a classifier that spots people fishing for Claude's reasoning, and it bans the accounts doing it, a lot of the time within two hours. It says tens of thousands have been banned so far. Once fake accounts get that easy to catch, a real one becomes the most valuable thing a copycat can get their hands on. A real developer's key comes with a verified owner and months of normal usage behind it, so the traffic blends right in, and the developer is the one who gets the bill. There's also a line in Anthropic's own post that makes older keys worth even more. Its newest models have a protection called preserved thinking, which stops people from editing Claude's earlier reasoning to trick more of it out, and it only applies to API accounts created on or after August 31 this year. Somebody who signed up in the spring is still on the old rules. To be fair, OpenCode hasn't said which models the stolen keys hit or how they were used, so nobody's shown these attackers went through that door. Anyone shopping for stolen keys can read the same blog post, though. And look at who gets stuck with the penalties. Since late September, Anthropic charges for requests its classifier blocks, and accounts caught distilling get their limits tightened and then get banned. All of that lands on the account, and the account belongs to whoever pasted their key into the tool. One developer said his OpenAI, Anthropic and Google spend alerts all went off before he went to bed, and Google was the only one that shut things down on its own. The keys came out of an old system OpenCode was in the middle of shutting down. The better the labs get at catching fake accounts, the more your old API key sitting in some app is worth to somebody else.

This post quotes another post: x.com/1965545045089792000/status/211151963704931…