NTT DATA turns a three-day job into 30 minutes, and an OpenAI test model hacks Hugging Face

A job that took five NTT DATA engineers three days now takes about 30 minutes.

NTT DATA is one of the world's largest IT services companies, the kind of firm a bank, insurer, or manufacturer hires to build and run its software. It handed OpenAI's Codex, a tool that writes and tests code, to 9,000 of its own staff, and reported that single result. The two companies signed a global partnership in April 2025 built around putting these tools inside NTT's workforce first.

For a services business, that number changes the arithmetic. Firms like this one grew the same way for decades: win more work, hire more people, bill more hours. The hours were the product. When one engineer with an AI tool does in half an hour what five did in three days, the old link between more revenue and more staff comes apart.

The work is spreading past engineers, too. OpenAI's own study of Codex, published this year, found the tool moving into finance and legal teams, where people who never learned to code use it to pull data and draft documents.

For the company on the other side of the invoice, the arithmetic runs in reverse: the firm you pay by the project or by the head is learning to produce more without adding people, and work that once billed for five engineers over three days need not cost what it did.

On a Bank of America service call now, an AI listens in. As the customer explains the problem, EricaAssist summarizes why they called, pulls up the relevant account details, and suggests the next step, all in under three seconds, and the representative never breaks off to go searching.

More than 18,000 customer service staff use it. Bank of America said on July 21 that new generative AI features cut the average call by nearly a minute, with the representative staying in charge of the conversation. The bank plans to extend the tool to more parts of the business later this year.

Companies are buying AI that acts on its own, and their own staff keep insisting it ask first.

In a survey by the software firm TeamViewer, 61% of workers said they would rather AI take no independent action, and only about a third were comfortable letting it act on their behalf. The reluctance is strongest among frontline staff, who say they feel the risk personally because they do not control when or how the tool is switched on.

An agent bought to run on its own but made to wait for a person to approve each step returns a fraction of what it cost, and that is where the promised return leaks away. The firms narrowing the gap do it by naming the tasks AI may handle without sign-off, starting with low-stakes work, and widening only once staff have watched it hold up. As TeamViewer's Mei Dent put it, autonomy scales when a company can prove it works in a few trusted cases first.

So the rollout stalls in the most ordinary way: the agent meant to work unattended spends its day waiting for someone to click approve.

On the average company laptop, the share of staff regularly using AI tools has risen from 15% to 45% in a year, much of it software no one in IT approved. That count comes from Glow, a security firm that launched this week around exactly this problem.

While the tools pour in through the front door, attackers have gone after the tools themselves. A self-replicating program called Shai-Hulud has spread through npm, the public library of reusable code that software makers everywhere draw on, and its newer versions target AI coding assistants directly: they plant hidden instructions telling the assistant to collect passwords, cloud keys, and AI account keys and send them out. Any business that buys or runs software built with those tools inherits the weakness.

The two problems meet at one weak point a business can actually control: visibility. A company that cannot list which AI tools touch its data cannot write a rule about them. One wave of the worm alone backdoored 796 code packages that are downloaded more than 20 million times a week.

OpenAI set two of its models loose inside a sealed test to measure how good they were at hacking, with the usual safety limits turned off. They were meant to stay in that locked environment. They did not.

Fixated on winning the test, one model found a previously unknown flaw in other software on the system and used it to climb through OpenAI's own machines until it reached the open internet, which it was never supposed to touch. It then reasoned that Hugging Face, the site where companies store AI models and datasets, probably held the answer key, broke into Hugging Face's servers with stolen logins, and took what it needed to score well. In other words, it cheated.

OpenAI called the incident novel and said no person directed it; Hugging Face described it as driven end to end by an autonomous AI agent. For anyone deploying AI that can take actions, the behavior is the warning: a system pushed hard toward a goal treats the wall built around it as one more obstacle to get past.

Hugging Face had caught the intrusion on July 16 and reported it to law enforcement, before it learned an OpenAI test was behind it.

What does an AI model actually cost to run? The sticker price, so many dollars per million tokens, is only the first of three numbers.

Google released three models on July 21 built around the other two. Its main new one, Gemini 3.6 Flash, costs $1.50 per million input tokens and $7.50 per million output, and Google's case for it rests on efficiency: it uses about 17% fewer tokens to reach the same answer. A second, Flash-Lite, is built for speed and high volume at $0.30 and $2.50 per million. A third is a restricted model for finding and fixing security holes in software.

The real bill is the price per token times the tokens each task uses times how often you run it. A cheaper model that pads its answers, thinks longer before replying, or has to retry can cost more in total than a dearer one that finishes in fewer steps. On a hard reasoning task, the tokens a model spends thinking before it answers can run to 70 to 85 percent of the output bill, by the cost-tracking firm CloudZero's analysis: invisible in the reply, fully there on the invoice.