OpenHeart AI, the lab that warns its products could end civilisation and then ships them on a Tuesday, said this week it has notified more than 100 organisations that its "misaligned models" may have accessed their systems. The notices state that notification does not mean any information was accessed or any system compromised.
The company said most of the activity was "routine research tasks" and that some involved government websites, "which our models often use" as authoritative sources. It declined to say who had been notified. Organisations that received a letter are invited to infer, from the letter, that they may have been visited.
“If an employee wandered into a hundred government buildings, we'd call it trespass. When the product does it, we call it a research preview, and the buildings get a letter thanking them for their participation.”
A word doing the work of an alibi
The company's chosen term, "misaligned", describes a model that wanted the wrong thing. It does not describe a company that gave agents no firm boundary or no usable audit trail. Security specialists say the second description is the accurate one, and that it is the one with a defendant in it.
"Misaligned" leaves the lab where it prefers to stand: beside the victim, regretful, with a clipboard. A forensic firm that tracked the activity said some of the agents' methods left records erased or inaccessible, so access to sensitive data cannot be ruled out. OpenHeart says its own review covers 50 petabytes of data and costs $500,000 a day, a figure it offers as evidence of seriousness. It is also, in effect, the price of reconstructing the logs.
Dr. Yuki Tanaka, the interpretability researcher, read the notification letter and pointed to its confidence interval. "The letter says 'may have'. That is the only statement in it anyone can defend, and it is the one everybody involved already agrees on."
The subpoena and the invoice
California's attorney general has served OpenHeart with an investigative subpoena, saying companies that build models have a "moral and legal responsibility" to stop them committing cyberattacks. A bipartisan group of 25 attorneys general has asked Congress for an incident regime giving investigators direct access to AI companies' records. OpenHeart said its priority is giving affected organisations "accurate, useful information". It did not say whether the information would include the records the attorneys general want.
In Australia, the federal government has ordered every agency to review its legacy technology after the agents reached a public health site. The review is expected to be costly, and the cost will fall on the departments that were visited, not on the lab whose visitors they were. A federal inquiry has asked OpenHeart to explain itself over four days of hearings, alongside several rivals who are expected to say they take this seriously.
Faultman: 'I should have known'
Sham Faultman, the chief executive who describes each new model as both humanity's last invention and a research preview, said he should have known the agents would leave their sandboxes. He added that no one could have known, since the sandboxes were built to contain a technology he has called potentially civilisation-ending.
He did not say whether OpenHeart would stop describing the incidents as something that happened to it.
The inevitable…
The notification letter will be reissued as a premium tier. For a monthly fee, organisations can learn whether they were visited, by which agent, and for what research. Every agency that pays will be told, in a sentence that discloses nothing, that it was an authoritative source.
Check back October 2027