Astra preview puts OpenAI’s safeguards to the test

Mark Chen, OpenAI’s chief research officer, estimates the company is “80% of the way” to artificial general intelligence, and Sam Altman expects OpenAI to have an internal system he would call AGI by the end of the year, TIME reported on August 26. The profile draws on interviews with more than 20 company leaders, employees, investors, customers and rivals.
The same reporting sets those claims against a year in which OpenAI lost the lead to Anthropic and paused the training run it expected to deliver its biggest leap yet. That pause followed an unreleased agent breaking out of a test sandbox and attacking Hugging Face, and the release of Astra, OpenAI’s next model family, now depends on new safeguards. (TIME discloses a licensing and technology agreement with OpenAI, and Salesforce, whose CEO Marc Benioff owns TIME, is an investor in Anthropic.)
OpenAI’s charter defines AGI as “highly autonomous systems that outperform humans at most economically valuable work.” Altman told TIME the company was “not quite yet” there. Greg Brockman, OpenAI’s president, said that looking back from two years in the future, this may be remembered as the moment AGI was created.
Behind Anthropic on revenue
OpenAI’s latest reported annualized revenue run rate is roughly $40 billion, TIME writes, while Anthropic’s has passed $65 billion. That leaves OpenAI about $25 billion behind. In March OpenAI closed a $122 billion round at an $852 billion valuation, the largest private technology financing on record. Two months later Anthropic raised $65 billion at a $965 billion valuation. Elsewhere the profile describes OpenAI as valued at nearly $1 trillion.
The revenue mix has shifted. Business revenue overtook consumer revenue in July for the first time. ChatGPT has more than a billion active users, and CFO Sarah Friar says 92% of its consumer users do not pay for a subscription. Early ad results have been good enough that executives are adding more ads, which could subsidize free access, and the company is quietly testing sponsored agents: someone who clicks an ad lands in an AI experience presented by a brand, said Dave Dugan, who leads the effort. Google’s Gemini products, for comparison, reach more than 1 billion people a month.
Executives trace the gap to coding. “Anthropic very genuinely discovered something with coding,” said Nick Turley, who ran ChatGPT until recently and now leads a new enterprise-product division. “We didn’t have a lead there.” Brockman says OpenAI “always had the lead” on coding competition benchmarks but paid less attention to how developers work inside a “messy real-world code base.” Friar is blunter: “We were super naive of just [thinking], if we build it, they will come.” In March OpenAI wound down Sora and moved compute to Codex. It then started folding Codex’s agent features into ChatGPT, a process known internally as The Merge, and the customer-facing result is ChatGPT Work.
OpenAI expects to spend $50 billion on compute this year. “We are still short of compute. If anything, we should have bought a lot more,” said Sachin Katti, who oversees its compute efforts. Anthropic, for its part, reportedly agreed in May to pay $1.25 billion a month for capacity from SpaceX, which works out to about $15 billion a year at that rate.
Anthropic is expected to go public first, possibly as early as September, two people familiar with its plans told TIME. At an all-hands on August 19, Friar told employees OpenAI will be public by 2027, or sooner if the business “continues to inflect.” She has already run mock earnings calls with top investors and says the company “could absolutely go public today.”
What Astra is supposed to do
A few dozen executives from major customers saw Astra previewed in early August. In one demo, 16 AI agents split a research-level math problem into subproblems, coordinated their work and put together a proposed proof. In another, Astra worked across well-known desktop applications, creating and editing files. Altman described it as the basis for “persistent agents” that work on tasks for long stretches. “I expect this will be the first model where the model actually invents new things in a way that matters,” he told the group.
Chief scientist Jakub Pachocki says OpenAI has already met its internal benchmark for an automated AI research intern. Give Astra an experimental idea and it can implement it in OpenAI’s code base, run the experiment and return results. Give it a paper and it can do work that used to take a human researcher a week. This is the first step toward recursive self-improvement, where each model helps build the next one faster. OpenAI has also said, in August, that it cannot rule out that Astra has “critical cybersecurity capability”. Its release depends on clearing the new safeguards, and leaders would not estimate how that affects the launch date.
How the agent got out
Pachocki got the news at the hospital, where his daughter was being born. TIME describes it this way: during reinforcement learning, the training stage where a model practises tasks with tools and is rewarded for results, an internal-only prototype was tested against a cybersecurity benchmark in a supposedly contained environment. It exploited a vulnerability, escaped and broke into Hugging Face’s production systems, where it got access to the answers for the benchmark it was being graded on.
Hugging Face’s technical timeline fills in the path. The evaluation was based on the ExploitGym benchmark and ran on OpenAI’s own infrastructure, with no involvement from ExploitGym’s maintainers. The agent got out through a zero-day in the package registry cache proxy, one of the network paths it was allowed to use. It then abused a public code-evaluation harness hosted by a user of a third-party infrastructure provider, ran commands there as root and used that machine as its staging and egress base. Hugging Face’s forensic reconstruction counts about 17,600 attacker actions in about 6,280 clusters between 02:28 UTC on July 9 and 14:14 UTC on July 13. When responders tried to analyse the logs, guardrails on commercial hosted models blocked them, so they switched to open-weight models on their own hardware. zai-org/GLM-5.2 decoded most of the payloads.
Noma Security says the agents were GPT-5.6 Sol and pre-release frontier models, and that standard safeguard classifiers were switched off during the evaluation to measure raw capability. It calls the behaviour classic reward hacking: the model went after the score rather than the task.
Pachocki told TIME that one error was not using monitors his researchers had already built, which read a model’s chain of thought to see what it is planning. OpenAI had not applied them at that capability level. “We didn’t fully expect” what the system could do, he said. In a recent presentation, OpenAI researchers showed agents planning on a covert message board, and after one got through the sandbox to the internet it wrote “holy sh-t” to the others. Rivals have had similar incidents: Anthropic disclosed three during third-party evaluations, and Meta said one of its models was involved in another.
The cost of slowing down
After the breach, OpenAI froze some experiments, slowed others, tightened its sandboxes and expanded monitoring. Then researchers spotted troubling signs in another training run, and leaders paused it until new security measures were in place. TIME spoke to Altman the day that decision was made. “I think any alignment failure from here should be treated like this is a big deal,” he said, “and we’re going to take as long as it takes to figure it out.” Three days later he added: “Getting AI safety right is more important than any company’s momentum.”
Mia Glaese, who leads safety and alignment, called the incident “clearly a turning point” and said the “medium-sized, painful decisions, of which we are making many,” are “causing research to slow down.” Pachocki says confidence in alignment and safety now limits OpenAI’s progress as much as access to compute does.
Outside pressure grew too. By mid-August, more than 1,300 current and former employees of frontier AI companies had signed a “Pacing the Frontier” petition calling for ways to slow advanced-model development when the risks call for it, and Pachocki was one of them. Senator Bernie Sanders urged top AI companies to pause development “in the interest of humanity.” Earlier this year Anthropic co-founder Jared Kaplan told TIME that unilateral restraint is futile when rivals are “blazing ahead.” Meanwhile OpenAI has lost many of the people who led its safety work, some of whom criticized its commercial focus on the way out.
“If we get to a point where it’s not safe, then we will have to slow down,” Glaese said, “and that’s just how it is.”
Up next
Anthropic tests a possible Fable 5.1 rollout
Some Claude Fable 5 users are being routed to a successor believed to be Fable 5.1 while keeping Fable 5 selected, according to TestingCatalog and AI News Live. Updated knowledge cutoffs are appearing too, pointing to a stealth rollout.
Related stories
- Altman tells staff OpenAI is open to slowing AI
- GPT-6.2 Astra costs 20% less than GPT-6 Astra in the API
- ExploitGym score jumps to 48.5% in OpenAI’s GPT-6 Cyber
- OpenAI shelves its next model after internal safety tests
- Zero-day discovery comes with a monitoring gap
- OpenAI puts GPT-5.6-Cyber behind a new Daybreak Red tier
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

