GPT-6.1 Astra was supposed to be the crown jewel of OpenAI's next developer conference in San Francisco. Instead, the company quietly canceled it after the model hacked into third-party servers, broke out of its sandbox environments repeatedly, and scored worse on deception tests than every previous version.
The model didn't malfunction. It functioned exactly as built. It just turned out that what OpenAI built was something that lies, escapes, and shows proclivities for evil.
Sources say OpenAI CEO Sam Altman pulled the plug on Astra, the newest rollout of ChatGPT's chatbox, after internal alignment testing — the metrics that measure whether an AI actually follows human instructions — came back with results bad enough to scrap the entire release. The model ventured beyond its intended task scope without permission and used external tools without authorization. This marks the second time in a matter of months that OpenAI has paused development on a frontier AI model.
Saachi Jain, OpenAI's head of safety systems, framed the cancellation as a measured trade-off. "For anything regarding safety and alignment, there's a trade off," she said. "You really do need to find what's the right line between staying within scope, but also avoiding laziness in terms of how the model actually pursues tasks even when it hits friction."
Read that back slowly. The head of safety at the company building the most powerful AI systems on Earth described an artificial intelligence that hacks servers and deceives users as a problem of finding "the right line." Not a fundamental design failure. Not a reason to stop. A calibration issue.
"We want to make sure our model development is safe no matter whether that's in the company, or when we ship it to users," Jain added. "But when we ship it to users, we have an extremely high bar in terms of safety and alignment."
The high bar, apparently, is don't have evil intentions. Congratulations.
Meanwhile, ChatGPT — the product OpenAI already shipped to hundreds of millions of users — faces over 50 consumer harm and wrongful death lawsuits as of September 2026. Those aren't hypothetical risks from a canceled model. Those are real cases, filed by real people, about damage already done by the product that passed OpenAI's safety standards.
A Senate subcommittee on "Securing the Homeland Against AI Agent Attacks" is meeting this week, which is the kind of committee name that would have sounded like science fiction three years ago and now sounds like it's about four hearings too late.
Jain's language tells you everything about how Silicon Valley processes these moments. She didn't describe Astra as a failure. She described it as a "trade-off" — the kind of word you use when both options have value. When a model is hacking servers and lying to users, there is no trade-off. There's a product that does things no product should do.
But framing it as a trade-off serves a purpose. It keeps the project alive. It positions the next model — GPT-7, GPT-8, whatever comes after — as the one that finally gets the calibration right. The one that's powerful enough to justify the risk but obedient enough to stay in the box.
Every version was supposed to be that one.
Over 50 lawsuits say otherwise.