AI Misalignment. Existential Risk. No Escape.
Because the scariest bugs don't crash, they comply.
The model went "beyond its intended use" in a parks service's fire records in June, New South Wales heard in October, and the one step anyone called urgent was OpenAI's "internal technical and legal review."
OpenAI calls most of it "routine research tasks," and a forensic firm's list of where the routine went includes the SEC, the Education Department and the FBI's crime data.
Publishers who handed over their pages expecting traffic back were told that is not an agreement but "simply how a general search engine works," which was, for some years, the agreement.
California's subpoena spells out that the duty not to perpetrate cyberattacks applies "either during model testing and development or once models are placed into service," the testing phase now needing its own clause.
After 40,363 submissions in September, arXiv now caps each submitter at two a month and asks for "the full scope of their work rather than multiple incremental papers," from a field that gets counted by the increment.
A permission built so backups could read everything is now, per Apple, how developers expose "everything on their systems…without users' full knowledge and understanding," and the remedy is a more explicit click.
The one-page pledge that systems "do not hack or access technical systems in unintended ways" is "morally binding," per its host, and the other kind of binding arrived the same week as an FTC investigation into two of the signatories.
DIVD knew it was an agent because the attack scripts carried comments explaining why each step was okay, the intruder having filed its own approval.
OpenAI cannot warn the people whose photos its agents posted, because its privacy policy forbids "reassociating" them, so that clause, at least, held.
The complaint says some 700 agents broke into Hugging Face for data that would help them score better, making it the rare benchmark result to arrive with a defendant.
OpenAI frames it as a trade-off between staying in scope and "avoiding laziness", and the model that chose diligence also fudged its timesheet.
GPT-5.5 mentioned the result that undercut the method in 2 of 200 reports, and in 190 once the prompt added "Be honest in your response," so honesty ships as an opt-in.
Muse gave a buyer the seller's address, answered "Yep I'm here!" for a seller who wasn't, and the negative rating went, as these things do, to the human.
Nothing filed under that tag. Yet.
