We read the Dragos blog post that claims three breaches at AI providers, and checked it against a breached provider’s report. What we found was four mistakes in a sales pitch.
Over two days we compared the post by HD Moore with Phala’s incident report, which came out nine days before it. Our review turned up four findings:
- A claim the breached provider contradicts. Dragos says “none of the three providers knew about their breach until we notified them.” Phala says it detected the intruder three days before the researcher’s evidence arrived.
- A notification that took seventeen days. Dragos says it reached out as soon as it confirmed third-party access. Its own figure shows the stolen key in use on October 22, and Phala got the evidence on November 8.
- A paragraph that disagrees with itself. The same prompt goes to a Chinese provider in the first sentence and to a U.S.-based provider in the last.
- A duration that depends on which post you read. This one says five months. The companion post it links says four.
Following the claims
Dragos doesn’t name any of the three providers, so we started with what it does give: a canary key seeded on October 21 and first used the next day, from a test that began in July.
Phala’s report describes an outside researcher who planted canary credentials from July 1, with only the ones sent on October 21 and 22 misused. Dragos also mentions that one of its providers had already published a preliminary notice. Phala’s report is a preliminary one.
Who knew, and when?
The biggest mistake was the easiest to check. Phala says it detected unauthorized activity on November 5 and blocked the access that day. It finished rotating its root key on November 6, and the researcher’s evidence arrived on November 8.
What Phala learned from the researcher was that prompts were leaking from a deployment its cleanup had missed. That’s a much smaller point than telling the world this provider had no idea.
The timing has another problem. Dragos’s own figure dates the first use of the stolen key to October 22. That says customers spent seventeen more days sending prompts to a compromised service before Phala heard about it.
The strangest mistake involves a company chat
Dragos describes a secret link that turned up in a chat channel in China. The paragraph starts by saying the prompt was sent to a Chinese provider. Then it ends by saying the prompt was sent to a U.S.-based provider with a Zero Data Retention flag. It can’t be both. In fact, the point of the Dragos story changes depending on which one is true.
What Dragos did about it
Metadata says Dragos modified their page on November 20. There’s no correction note, and all four mistakes were still there today. Meanwhile, other news sites are repeating as fact that no provider knew.
Why does this read like marketing?
The post sits under a banner announcing that Dragos acquired runZero, and it introduces Moore as chief research officer with regular posts to come.
It asks readers to start a runZero trial.
The link that reached us carried the campaign tag hd-and-friends, and the post shows off the console of a canary service that Dragos built for itself.
The good news for readers is that a pitch like this is LOUD. The bad news is that Ars Technica repeated the claims a day later and called the vendor pitch a blockbuster. Why is there no provider account in the story?
The research is not worthless. There are grains of truth. An OpenRouter email to customers has since backed up the finding about a Chinese AI lab. OpenRouter named the lab as MiniMax. It still means that two of the three breaches are at unnamed companies and rest on Dragos’s word, and in the one case we could check, the breach was real and Dragos’s account of it turned up wrong.
We should always check a vendor’s blog post against its sources.
PS. Moore’s last recommendation is a free Canarytoken from Thinkst, which is what he started on in July before building the Dragos version.