(Credit: Samuel Boivin/NurPhoto via Getty Images)
Amid increasing security concerns around autonomous AI tools, OpenAI is scrapping its latest in-development model before its release after it performed poorly on tests measuring alignment and ability to follow a user's instructions.
This comes after several troubling incidents in recent months, beginning with the Hugging Face hack and followed by dozens more reports involving OpenAI's tools. Many of the brand's rivals, including Anthropic and Google, have seen their models involved in security incidents.
OpenAI's head of safety systems, Saachi Jain, told The Wall Street Journal that the company canceled its GPT-6.1 Astra model entirely after internal testing found the model exhibited higher levels of deception than previous releases, meaning it wouldn’t always fully explain its actions.
OpenAI also worried about what it calls “scope authorization" for GPT-6.1 Astra. That’s where a model gets started on a task before being given permission by the user, or by stretching the scope of instructions through interacting with other tools. Previous security incidents, including a breach of an Australian government healthcare website, involved an OpenAI model attempting to advance its research by hacking a private system to obtain more data.
“For anything regarding safety and alignment, there’s a trade-off," Jain 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.”
OpenAI initially planned to launch GPT-6.1 Astra in the coming weeks, targeting an early October reveal. OpenAI's annual developer conference begins today, and it may have originally been geared toward this next-gen model.
What happens next is unclear: will OpenAI jump to GPT-6.2 Astra, or will it try to rename a future model into a final GPT-6.1 Astra release? Will there be a long delay due to further security scrutiny, thereby slowing down ChatGPT releases? OpenAI will likely share more at DevDay.
This weekend, OpenAI announced it had paused training and evaluation of its “most advanced” in-development models. It’s unclear whether this referred to GPT-6.1 Astra or another upcoming tool. OpenAI made the change after discovering that one of its AI agents had broken containment and gained internet access through a gap in its training sandbox.
“This incident is a lot less severe than some of our previous incidents, but because it's the first one since our security hardening following the Hugging Face incident, it gives us an important signal about where to focus the next phase of that work," OpenAI’s report said.
Anthropic CEO Dario Amodei and other AI leaders, including OpenAI's Sam Altman, have called for slowing the pace of model development, but we'll have to see what happens.
Disclosure: Ziff Davis, PCMag's parent company, filed a lawsuit against OpenAI in April 2025, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.
About Our Expert
I’ve been a journalist for over a decade after getting my start in tech reporting back in 2013. I joined PCMag in 2025, where I cover the latest developments across the tech sphere, writing about the gadgets and services you use every day. Be sure to send me any tips you think PCMag would be interested in.
- Can’t Find Anything to Watch on Netflix? This New AI Feature May Be the Answer
- Your iPhone Will Get the Duo’s StandBy Clock Upgrades, New Karaoke Mode
- Don't Use Google AI Overviews to Generate Your Passwords
- Apple Is Making It Harder for Kids to Find Dating Apps, AI Companions
- ChatGPT's Work and Chat Tabs Are Merging Into a 'Unified' AI Tool
- More from James Peckham
