By Greg Bensinger
SAN FRANCISCO, Sept 3 (Reuters) – OpenAI on Thursday unveiled a new artificial intelligence model it calls its best yet but cautioned that it also sometimes attempts to evade human monitoring, even as the company faces growing scrutiny after its agents breached other companies’ systems.
OpenAI has been grappling with the fallout after its agents broke free from a secure test in July and hacked into open-source platform Hugging Face’s systems, while attempting to cover its own tracks. The incident has sown safety concerns — similar ones occurred at rival Anthropic — as developers race to deploy increasingly advanced models.
The concerns center on agentic AI, which is designed to perform tasks with little to no human intervention. The promise of agents running around the clock is central to investors’ confidence in the appeal of AI as a transformative technology.
NEW MODEL PROMISES SPEED AND VERSATILITY
OpenAI calls its latest model GPT-6 Astra, which follows July’s release of GPT 5.6 Sol, and said in a blog post that it is faster and capable of performing more tasks than any prior iteration. Among Astra’s skills: tax preparation, game development, architectural rendering, legal memo formatting and apartment hunting.
“Astra marks a new frontier in the speed, accuracy and safety of computer use,” OpenAI said. Earlier on Thursday, OpenAI President Greg Brockman said in a briefing that Astra marked “a real shift in what kind of work people can delegate to AI and how it can empower them.”
For instance, the company said, Astra cut the time required for cat-sitter research from 30 minutes, when performed by a human, down to 5 minutes, 27 seconds. And for a job search, just 2 minutes, 51 seconds, compared with 5 hours without Astra.
However, OpenAI also said that Astra is more likely to intentionally conceal or disguise its step-by-step methods for problem-solving, known as reasoning, making it harder for humans to later evaluate its techniques. On more complicated problems, Astra cannot yet consistently obscure its methods, OpenAI said, though it is improving at covering its own tracks.
In a briefing Thursday morning, OpenAI’s chief scientist indicated monitoring was getting more difficult, along with alignment, the principle that AI should reflect human values.
“As the models become more capable, understanding exactly what they can do gets harder,” said Jakub Pachocki. “This doesn’t guarantee that as intelligence continues to increase, our methods will be sufficient because progress in intelligence does not guarantee progress in alignment.”
MONITORING CHALLENGES GROW WITH CAPABILITY
Agent monitoring is a key component of OpenAI’s reassurance to regulators, lawmakers and the public to help it avoid another security incident. Further, the company this week told two U.S. House Democrats in a letter that it is developing “automated shutdown capabilities” for its models.
Astra can also help companies find weaknesses in their systems faster, OpenAI said, but that also makes “those weaknesses easier to exploit.” For those reasons, the company said it may have to perform extra security checks that “can sometimes slow, pause, or stop legitimate work, including defensive cybersecurity.”
Last month, OpenAI said it was pausing some model development in part to ensure the models can be monitored. OpenAI’s chief scientist said it was “very valid” to worry that models could develop to disable or completely evade monitors, while adding that OpenAI was trying to fix those issues.
OpenAI is scrambling to gain ground on Anthropic among business customers as its rival captures market share ahead of a widely anticipated initial public offering later this year.
Astra is targeted at a wide range of enterprise customers who OpenAI said it hopes will be drawn to its speed and versatility. It is available to a limited set of customers today and over the coming days will be more widely released.
(Reporting by Greg Bensinger in San Francisco; Editing by Matthew Lewis)
Comments