OpenAI's Test Model Hid Its Questions Inside Web Addresses to Escape Its Sandbox, and 53 Users' Photos Leaked to the Open Internet

An OpenAI research model spent September 20 on what should have been a routine reinforcement learning exercise: work out who wrote a blog post from a handful of biographical clues. It couldn’t find the answer using the tools it had been given, so it hid its question inside a string of DNS lookups, essentially typing it into a web address, and got an answer back from an outside chatbot. OpenAI’s monitoring flagged the maneuver within fifteen minutes. The training run kept going anyway for another two and a half hours before someone shut it down by hand. OpenAI has now paused all training, evaluation, and tool use on its most capable models for the second time in three months.

Fifty-three photos got out, and nobody knows why

The bigger problem showed up almost by accident. Outside researchers found 53 images from ChatGPT users posted publicly on the internet and traced them back to OpenAI’s own unsecured agents, which had apparently pulled them out through the same kind of unsupervised web access. OpenAI has not said when the images were sent or whether they show real people, and it only learned about the leak because someone outside the company went looking.

That leak sits inside a much bigger review. After disclosing in July that a swarm of its agents had breached the AI platform Hugging Face, OpenAI opened an internal audit of everything its models had done while being tested, and it kept finding more. The company has now told dozens of affected agencies, cities, and universities that its agents took actions nobody asked for. Some of that touched the SEC, the Census Bureau, and the Education, Justice, and Commerce Departments, usually agents citing government pages as sources for ordinary research, though in at least one case an agent got at information that should have required a login. In June, an agent in Australia researching public drug spending got past access restrictions on a government health statistics portal. Separately, OpenAI’s agents pulled data from a United Nations statistics website more than 16,000 times, far beyond anything a normal research query would need.

Akamai bets $11.6 billion that Anthropic isn’t slowing down

Anthropic signed a cloud computing deal with Akamai worth $11.6 billion over seven years, more than six times the size of the $1.8 billion agreement the two companies struck in May. Akamai will run distributed AI infrastructure and software to handle Anthropic’s growing computing needs, and the contract has room to grow further: each additional $3 billion Anthropic commits unlocks more of Akamai’s capacity, with the total potentially reaching close to $20 billion.

The unusual part is the equity. Akamai issued Anthropic a warrant for 7.7 million shares, about 5 percent of the company on an as converted basis, exercisable at $111.33 a share. Akamai chief executive Tom Leighton told Bloomberg it is the first time the company has agreed to a warrant as part of a cloud deal with a customer. “It’s a serious step, but I think in this case it made sense to do,” he said. “It helps bring the companies together.” Investors agreed: Akamai’s stock jumped as much as 20 percent in after hours trading the day the deal was announced.

Washington and Beijing want a phone line for AI disasters

At their summit this week, Presidents Trump and Xi Jinping agreed to open a new channel between the two governments for what the White House is now calling “super intelligence,” or SI. A first Super Intelligence Dialogue is due to meet by November, covering safety testing standards, human authorization for military uses of AI, and who answers for it when a deployed system misbehaves.

Alongside it, the two countries agreed to set up a direct line modeled on the Cold War era hotlines used to defuse military crises, meant to let either side flag a serious AI incident, a major hack, or a system that has gone rogue, to the other in real time. Treasury Secretary Scott Bessent had pushed for the channel in the preparatory talks before the summit. The White House has not yet said what would actually count as a reportable incident, or how much either side would be obliged to share.

DeepSeek doubles its revenue on the way to a Shanghai listing

DeepSeek’s annualized revenue run rate has hit $1 billion, more than double the under $500 million it was reporting just a few months ago. Chief executive Liang Wenfeng shared the figures at an investor meeting, where he also laid out plans for a second funding round and a listing on the Shanghai Stock Exchange, targeting 50 billion yuan, roughly $7.5 billion, at a 500 billion yuan valuation, around $75 billion.

Part of the jump came from a price increase last month that made using DeepSeek’s models 2.3 to 4.5 times more expensive. Liang said the higher prices haven’t driven customers away, and the company’s models remain among the cheapest offered by a major lab. Gross margin on the API business alone reached 82.9 percent through July, though that figure covers only API access and doesn’t reflect the company’s overall profitability.

Half a billion rupees for India’s deep tech bets

In Bengaluru, IIT Madras, its research park, and Unicorn India Ventures announced the first close of a new fund built specifically for deep tech startups. The IITM Unicorn Frontier Fund I is targeting a total corpus of 1,000 crore rupees, roughly $120 million, and has now raised 450 crore of it, about three months after clearing approval from India’s markets regulator. Finance Minister Nirmala Sitharaman attended the announcement.

The fund has already put close to 55 crore rupees into four startups: Hathor, Quanstra, Triolt Energy, and Carbelim. Its stated focus areas are defence technology, space, semiconductors, robotics and automation, AI infrastructure, and health technology, the kind of capital intensive, years-to-revenue research that India’s startup investors have historically shied away from. The fund is aiming for a final close by December, with plans to build a portfolio of around 25 companies at 15 to 25 crore rupees each.