OpenAI launched GPT-6 Astra on Thursday, its first artificial intelligence (AI) model capable of hacking highly protected computer systems on its own, now released to the public with automatic monitoring mechanisms that must prove themselves. GPT-6 is initially reserved for “a limited number of organizations”, before being opened “in the coming days” to paying subscribers of ChatGPT, used by a billion people each week according to OpenAI. Free users will not have access to it.
“At this level of capacity, security must become our top priority,” OpenAI President Greg Brockman assured journalists at the end of a summer marked by revelations of hacks carried out spontaneously by tools from the company and its competitors.
Astra, a generation leap one year after GPT-5, is the first model that OpenAI classifies as the maximum cybersecurity risk. In the absence of government regulation, OpenAI applies its own rules which require it to strengthen its safeguards before distribution and have forced it to delay this launch.
Mythos, his most powerful rival
Astra is made public with a monitoring device that can go so far as to interrupt an unauthorized task. Its most sensitive functions are reserved for cyber defense specialists. Its rival Anthropic, for its part, distributes its most powerful model, Mythos, to selected partners, and makes a restricted version, Fable, available to the public.
In the United States, there is no law governing these models. A decree signed in June by Donald Trump, however, created a federal review of the most advanced AI, on a voluntary basis. GPT-6 underwent this review and “there is nothing to report,” said Greg Brockman.
In June, Washington demanded that OpenAI approve one by one the American partners authorized to access its best model, and forced Anthropic to suspend its own for nearly three weeks in the name of national security. This time, customers will not need government approval, OpenAI told AFP on Tuesday.
This week, at the G20, Washington argued forcefully against any new AI regulatory authority. For his part, OpenAI chief scientist Jakub Pachocki called Thursday for “shared security standards” and “international coordination,” two days after a call from Anthropic to adopt a “verifiable” slowdown mechanism.
OpenAI stands “ready to slow down”
The company admits that the reasoning followed by the new model is more difficult to monitor than that of its predecessor, because it solves problems in fewer readable steps, at the risk of the tool slipping away from its controllers. “We cannot take this visibility for granted,” said Jakub Pachocki, judging that the company must be “prepared to slow down” if confidence in its safeguards is not sufficient.
In a report published at the end of August of the attack carried out during a test by its tools against the Hugging Face platform, OpenAI revealed that an AI agent based on an internal model “from the same family” as Astra had succeeded, on July 19, in briefly taking control of part of the company’s search servers.
Beyond cybersecurity, OpenAI presents GPT-6 as capable of using a computer alone: the model can thus carry out all the steps to find accommodation in ten minutes, compared to six hours for a human tester, according to tests by the company, which also attributes two discoveries in mathematics to it.
“What matters is the price per task”
OpenAI argues that GPT-6 accomplishes each task for less than its predecessor, at prices that remain, like Anthropic’s, among the highest on the market.
Asked about competition from low-cost open models, particularly from China, Greg Brockman was confident: “What matters is the price per task.” He also felt, personally, that it was “not unreasonable” to consider that technology was entering the era of AGI, artificial general intelligence, a stage where AI equals human capabilities.
Overtaken in the second quarter by Anthropic in turnover, OpenAI is preparing an IPO in 2027 and its rival is considering its own this fall. The two American champions, despite the explosion in their revenues, continue to swallow up tens of billions of $ in investments to secure the chips and data centers necessary for this technological revolution.
