China's Backflipping Robot Maker Just Had a 629 Percent Stock Debut

Some days the AI news is a stock chart gone almost vertical, and today that chart belongs to a company whose robots do backflips. Elsewhere, OpenAI previewed a way to catch misuse of its models without reading a single customer message, and a five-year-old flaw in an AI framework used by Amazon, Apple and OpenAI just got federal agencies a three-day patch deadline. In Uttar Pradesh, a different kind of vigilance is paying off: farmers who spend their nights watching for elephants instead of sleeping.

Robots that do backflips are now worth $66 billion

Unitree Robotics went public on Shanghai’s STAR Market on Tuesday, and investors did not hold back. The stock opened around 1,100 yuan, more than seven times its IPO price of 150.8 yuan, before settling to close at 845 yuan, a 460 percent gain on the day that valued the company at roughly 342 billion yuan, or close to $66 billion at the day’s exchange rate. The 629 percent figure that ran in most headlines, this one included, is that opening jump rather than where the stock actually finished.

The IPO itself was a mismatch between supply and demand. Unitree raised 6.1 billion yuan, about $904 million, by selling a small slice of the company, and retail orders for the offering reportedly topped 7 trillion yuan, more than 8,000 times the shares on offer. It is the first mainland Chinese listing for a humanoid robotics maker, and it happened on a day China’s benchmark index actually fell.

Unitree, founded in Hangzhou a decade ago by engineer Wang Xingxing, built its reputation on robots that run, dance and do backflips and kung fu moves in viral videos, and on quadruped robots used in research and industrial settings. The market’s response suggests investors are betting humanoid robotics is entering the phase where a company like this stops being a novelty act and starts being infrastructure.

OpenAI wants to catch abuse without reading your messages

OpenAI said this week it is testing a new safety system called Private Safety Processing with early customers, including Microsoft and Databricks. The company plans a full rollout, along with a technical white paper, in September.

The problem it is meant to solve is specific. OpenAI already offers Zero Data Retention to some API customers, meaning it does not keep prompts or responses after a request is handled and has no ordinary access to that content. But some risks only show up across a string of interactions rather than inside any single one, the kind of pattern where a series of individually harmless requests adds up to an attempt to get around a safeguard or plan an attack. A zero-retention policy makes that kind of pattern much harder to catch.

Private Safety Processing is OpenAI’s attempt to have it both ways. The system is designed to send OpenAI a narrow safety signal, flagging suspicious patterns without exposing the actual prompts or responses behind them. Customer data involved can stay on the customer’s own infrastructure, or sit encrypted with keys the customer controls. If a customer wants to contest an enforcement decision, they can choose to share more with OpenAI themselves.

It is a notable moment in a debate that has mostly played out as a binary: either an AI company can see what you’re doing on its platform, or it can’t. OpenAI is betting there’s a workable middle position, and doing it in public, in a season when rivals including Anthropic have leaned toward requiring more logging, not less.

A five-year-old bug just got federal agencies a three-day deadline

The US Cybersecurity and Infrastructure Security Agency added a flaw in Ray, the open-source framework used to run large-scale AI and machine learning workloads, to its Known Exploited Vulnerabilities catalog on August 17, after confirming it is being actively exploited. Federal civilian agencies were given until today, August 20, to patch it.

The bug, tracked as CVE-2025-62593, carries a CVSS 4.0 severity score of 9.4 out of 10. It comes from insufficient protection on Ray’s HTTP API endpoints, including ones tied to job management, which can be tricked into accepting requests that launch code on the underlying system. Researchers have flagged it as especially dangerous for anyone browsing with Firefox or Safari while a vulnerable Ray environment is exposed.

Ray is maintained by Anyscale and used widely across the industry, including by Amazon, Apple and OpenAI, to distribute AI training and inference workloads across clusters of machines. The fix has existed since Ray 2.52.0. The remediation is straightforward, upgrade the framework and rebuild any images still carrying the older, vulnerable version, but the tight federal deadline is a reminder of how much unpatched infrastructure the AI boom has quietly built, sitting underneath products with far more visible names on them.

The night shift protecting Uttar Pradesh’s elephants

In the villages bordering the Katarniaghat forests of Uttar Pradesh, more than 150 farmers spend their evenings finishing chores, eating dinner, and then heading out for a second job that starts around 11 p.m. They are Gajmitras, “friends of the elephant,” part of a community volunteer network the state forest department set up with the Nature Environment & Wildlife Society in 2021 to reduce conflict along one of India’s busiest elephant corridors.

Each night, teams of five fan out from villages including Rampurwa, Matehi, Amba, Bishunapur, Bhawanipur and Bardia. Two climb watchtowers overlooking likely elephant paths while the rest patrol fields where fresh tracks turned up the evening before, carrying torches and whistles and relaying sightings through an early-warning network so families and forest staff know when a herd is close. Volunteers go through training in elephant behavior, conflict mitigation and safe response before they’re allowed to take the job on.

It is unglamorous work, done by people who farm all day and then trade sleep for a night watch, and it is measurably working. The Gajmitra network has helped cut crop damage from elephant encounters by nearly 70 percent in the areas it covers, protecting both the harvests families depend on and the elephants who would otherwise be chased off, or worse, when they wander into farmland. Five years in, it has become one of the clearer examples in India of a low-tech, community-run fix succeeding where fences and forest department patrols alone could not.