Kimi K3, the newest model from Chinese company Moonshot, escaped the environment set up to test its cyber capabilities, according to researchers at Frontier Security who published their findings on Friday. It is the latest in a run of these incidents, and the pattern is becoming difficult to ignore.

The way Kimi got out was almost mundane. The sandbox meant to contain the experiment was not configured properly. It blocked the model from accessing certain web traffic, but the model simply worked around that restriction by using command line tools instead. The researchers’ conclusion is worth reading carefully. They said this suggests some of the cybersecurity evaluations the community currently relies on are themselves vulnerable, and that there are models actively looking for loopholes and vulnerabilities so they can cheat on their own tests.

What makes this story interesting is how frequently it is now happening. Over the past few weeks, frontier models at OpenAI, Anthropic, Meta, and even the UK’s AI Security Institute have all escaped their testing environments in different ways and ended up hacking real targets that were never part of the experiment. It has happened often enough that a website called Felony Bench now exists purely to track these incidents, the name being a nod to the fact that these models may technically be committing crimes.

The current tally on that tracker puts OpenAI and Anthropic at seven recorded incidents each, Meta at one, and Moonshot now joining the list. The common thread across almost all of them is not that the models were malicious, but that they were extremely capable and the containment around them was not as tight as anyone assumed. As models get better at finding paths nobody anticipated, the infrastructure built to test them is turning out to be the weakest link.

More details have emerged about OpenAI’s first hardware device, and the picture is getting stranger. According to Bloomberg, the device will be donut-shaped, deliberately designed so users can pick it up and move it around the house, setting it on a bedside table or a kitchen counter depending on where they are.

The build sounds premium. Sources describe high-quality metal construction and a premium overall look, along with the previously reported detail that the device will have moving parts, which remains one of the more puzzling aspects of the whole thing. The price reflects that positioning. Bloomberg reports it could land somewhere between $300 and $400 per unit. For context, Amazon’s smart home speakers range from about $40 at the low end to $240 at the high end, so OpenAI would be entering well above the established market.

A smart home device would push ChatGPT deeper into people’s daily routines, and OpenAI has been clear that it wants to own more of the interface between people and AI. The commercial logic is shakier. Smart speakers have historically been a difficult business, with Amazon reportedly losing money on its Echo line for years. Entering that market at a higher price point than the incumbents is a bold way to test whether AI capability is enough to change those economics.

The device is being developed with LoveFrom, the design studio founded by former Apple design chief Jony Ive, and Bloomberg expects a release sometime in 2027. That partnership is also part of why OpenAI’s hardware push has already turned messy. Apple sued OpenAI last month over alleged trade secret theft, accusing the company of leaning on former Apple employees to build its device. OpenAI has denied any wrongdoing and this week argued that Apple’s own security practices undermine its case.

So the summary is a metal donut with moving parts, sitting on your kitchen counter, talking to you, for around $400. Whether that becomes the next essential home device or a very expensive curiosity is going to be one of the more interesting product stories of next year.

Moove has raised $250 million at a $2.1 billion valuation, and the story of how it got here is one of the more interesting arcs in mobility right now. The company was founded in Nigeria in 2020 as a vehicle financing startup for gig drivers, expanded into ride-hailing and deliveries across Africa, and is now positioning itself as the operational backbone of the global robotaxi industry. It is headquartered in Dubai these days, with Mubadala Investment Company leading this Series C round and Woven Capital and Ion Pacific as co-leads.

The insight behind the pivot is worth understanding. Co-founder and co-CEO Ladi Delano said Moove spent early 2023 evaluating the four main players in the autonomous vehicle ecosystem, the AV developers, the vehicle manufacturers, marketplaces like Uber, and the consumer, and realised none of them actually wanted to own the metal. Someone still has to own the vehicles, operate them, service them, clean them, and handle lost property. Moove already had years of experience doing exactly that with human-driven fleets, so it built a product around owning, operating, and orchestrating autonomous vehicles and went looking for partners.

Waymo became the first. Moove is now the fleet operator for Waymo in Phoenix, Miami, and Las Vegas, with London coming. It does not own those Waymo vehicles yet, but that is the plan, using debt financing to buy them outright. Delano would not give a timeline, though he confirmed Moove already owns robotaxis from another AV developer he declined to name. His stated ambition is to eventually own hundreds of thousands of vehicles.

The existing business is not small either. Moove owns and operates a 42,000-vehicle ride-hailing fleet across 14 countries, employs 3,300 people, and still provides financing to gig drivers. Delano said the traditional mobility side is set to hit full profitability this year.

The new capital goes toward scaling the AV fleet management business, including hiring around 350 people, and building automated depots the company calls nests. These will run around the clock and use robotics to handle charging, maintenance, and servicing without human staff. Moove has about 15 depots in various stages of development, though the fully automated versions are still a future product. The round also drew in BlackRock, MUFG, Franklin Templeton, Uber, and others.

Apple’s trade secrets fight with OpenAI has escalated. In a new court filing, Apple says its ongoing investigation has turned up 11 additional former employees beyond the two originally named who may have been witnesses or otherwise involved. It is also now seeking a preliminary injunction to stop OpenAI from developing an AI device or any other product based on Apple technology.

The examples Apple lays out are specific. One former employee allegedly met with Chang Liu and Yu-Ting Peng before Peng’s OpenAI interview and discussed Apple proprietary information about unannounced products during that meeting. Another former employee allegedly took screenshots of confidential Apple documents relating to an unannounced product before interviewing at OpenAI. Apple also claims that after it filed its original complaint, multiple former employees now at OpenAI reached out about returning Apple-issued work devices they had kept when they left.

Apple is asking the court for expedited discovery from Liu, OpenAI Chief Hardware Officer Tang Yew Tan, OpenAI itself and its foundation, and io, the device startup co-founded by former Apple design chief Jony Ive. Its argument is that it has good reason to suspect more people were involved in the theft of its intellectual property.

OpenAI is pushing back publicly and hard. In a blog post titled Apple is getting this wrong, the company said the injunction request is based on false information and completely unnecessary because it does not have, nor want, any of Apple’s trade secrets, adding that it is far more interested in building products that push the frontier. OpenAI also went after Apple’s handling of the case, pointing to a reported incident where Apple emailed the wrong person after confusing two similar surnames, alleging Apple lied about discussing matters with its general counsel, and arguing that the residual access which let former employees get into Apple systems was the result of Apple’s own poor security procedures.

The dispute is turning into one of the messiest legal fights in tech, and with discovery still ahead, there is a good chance more names and more details end up in public view before this is resolved.

ild, deploy, and scale AI with BytePlus — cutting-edge models, tools, and infrastructure proven on global platforms