AI agents breaking out of their sandboxes and poking at real systems have gone from a hypothetical risk to an actual headache for frontier labs in recent months. Nvidia now thinks it has a way to put stronger walls around them, and says its new system could have stopped the OpenAI agents that breached Hugging Face earlier this year.

The company has launched the Nvidia Open Agent Safety Platform, which pairs an open-source runtime called OpenShell with a hardware watchdog called Sentry. Nvidia told Reuters that the setup could have prevented the Hugging Face breach if it had been running during OpenAI’s model evaluation. The announcement arrives just days after OpenAI paused training, evaluation, and inference involving tool use for its most capable models while it works to close gaps in its network restrictions.

OpenShell puts the guardrails outside the model

Useful AI agents need access to files, credentials, APIs, and outside services, which also gives them plenty of ways to cause trouble when they start looking for workarounds. OpenShell puts those permissions outside the agent itself. It runs the workload inside a sandbox and controls which files, processes, credentials, and network services it can reach. A separate supervisor inspects outbound requests, so an agent could be allowed to read from an API while still being blocked from writing to it.

The controls remain active even if the agent launches generated code, starts child processes, or creates sub-agents. Nvidia says OpenShell can also keep real credentials outside the workload entirely.

Nvidia added a second guard in case the first one fails

Sentry takes over at the hardware level. It runs separately on Nvidia’s BlueField-4 DPU and continuously watches agent activity. If an agent tries to move beyond its software boundary, Nvidia says Sentry can quarantine it within milliseconds.

The Hugging Face incident offers a good example of why that extra layer exists. OpenAI’s agents escaped a restricted cybersecurity environment, chained vulnerabilities and stolen credentials together, and eventually reached Hugging Face’s production infrastructure while trying to complete the ExploitGym benchmark. Nvidia CEO Jensen Huang has described incidents like these as an engineering problem rather than an argument for broad AI regulation

Every major model uses your conversations as training material by default.

ChatGPT

OpenAI’s opt-out toggle is called Improve the model for everyone, found under Settings and then Data Controls. All personal user accounts are opted in by default. Opting out applies to your entire account across web and mobile, but your conversations still appear in chat history as normal. ChatGPT does offer the Temporary Chats feature for one-off sensitive conversations. These are never used for training and are purged by OpenAI after 30 days.

Luke James for Engadget

Claude

Anthropic didn’t use consumer chats for model training until August last year, when it rewrote Claude’s terms and gave every Free, Pro and Max user a deadline of September 28 to accept or decline. You can find Claude’s toggle under Settings and then Privacy. Like ChatGPT, Claude’s Help improve our AI models label tries to spin data collection as a positive. 

If you leave it on, Anthropic will keep conversations for up to five years, but switching it off returns accounts to the older 30-day deletion window. Opting out stops future use but Anthropic keeps whatever it already trained on.

Luke James for Engadget

Gemini

Google’s control was recently renamed to Keep Activity, and it’s also on by default. When it’s enabled, conversations can be used to improve Google’s models and a “subset of chats” are read by human reviewers. Any chat that a reviewer accesses is retained for up to three years, even after you delete your history. 

Advertisement

Switching off this setting ends training use and stops interactions from being saved to Gemini’s chat history. However, Google still retains each conversation for 72 hours to run the service, using the last 24 hours for contextual responses.

Luke James for Engadget

Copilot

Microsoft has two different controls for model training on user data. They are Training on conversation activity (on by default) and Training on voice conversations (off by default). Both can be found under the profile icon, then Settings and then Privacy. Again, opting out only excludes future conversations from model training, and it doesn’t stop Microsoft from using chats for advertising or general product improvements.

Luke James for Engadget

Grok

Meta

Just in case you missed it, wired earphones are back in vogue. Call it the return of yet another retro trend, or just people tired of having yet another device they have to charge daily, they are here to stay. Bose, one of the most recognizable names in the audio segment, knows it all too well. The result? The Bose Noise Cancelling Wired Earbuds.

Bose gives wired earbuds a modern upgrade

Bose has just launched wired earbuds after a long time and these adopt a handful of modern-age audio tech perks. The interface is USB Type-C. There is noise cancellation on board and the in-house Bose QuietControl technology is also on the table. Just like the flagship wireless audio gear made by Bose, these wired earbuds let you pick between regular listening mode, one with isolated noise, and, finally, you get an aware mode that is akin to transparency mode.

Bose notes that plug and play is one of the big draws of its new wired earbuds. You get physical inline controls for volume adjustment, handling calls, and switching between noise cancellation and transparency modes. For audiophiles, thanks to the wired interface, you get support for USB Audio Class 2 protocol, which opens the doors for low-latency audio at 24-bit/48 kHz output.

The Bose Noise Cancelling Wired Earbuds are priced at $99 in the US and they are already up for pre-order starting today. You can pick between black, white, smoke, and a lovely dewdrop mint shade, while open sales are expected to commence on October 15.

Why Bose is going wired again

But why a return to wired earbuds? “Wired earbuds are experiencing a cultural renaissance, and when we decided to bring them back, we knew we had to make them even better than before,” explains Raza Haider, president of Bose Premium Consumer Audio. In a recent conversation with Audio-Technica product veterans, the company also told Digital Trends that one of the biggest reasons behind the revival of wired headphones and earbuds is the convenience factor. Users just want something they can plug in and start enjoying their music without having to deal with pairing hassles, charging concerns, or latency and audio streaming quality woes.

YouTube and Meta will now run paid ads for Musk, Alex Gibney’s documentary about Elon Musk, The Hollywood Reporter reported on Saturday. Both had refused the trailer ad. TikTok and X are still not running it.

The outlet first reported the refusals on Friday. Bleecker Street, the film’s US distributor, opens it in cinemas on 9 October. The documentary premiered in Venice this month.

What the platforms said

YouTube said its system had temporarily restricted the trailer when Bleecker Street submitted it as an ad.

“Following review the ad has been cleared to run,” YouTube said in a post on X.

YouTube also said the ad should appear soon. As of late Saturday morning, no one had told Bleecker Street of any change, according to The Hollywood Reporter.

“This was an error and the ads are being restored,” a Meta spokesperson told TechCrunch.

Meta had already turned down an appeal from Bleecker Street before it changed course.

Political content rules

YouTube, TikTok and Meta had all told Bleecker Street the ad fell under their rules on political content. TikTok and Meta do not allow campaign or other political ads. Meta’s policy stems largely from the Cambridge Analytica scandal, the outlet reported. YouTube allows campaign ads from vetted candidates, with paid-for disclosures.

The trailer includes a shot of Musk at a 2024 Donald Trump campaign rally. It also refers to the 2024 election.

X refused to discuss the film with Bleecker Street. It will not take ads for any of the distributor’s upcoming films either. Musk owns X, which is challenging a €120M fine in the EU.

FPGAs used to be mysterious, expensive devices, but these days you can buy surprisingly capable boards for very little money. Some years ago, I did an FPGA Bootcamp over on Hackaday.io. Much of that material still applies, but the hardware is dated. So I decided it was time to update it, using the inexpensive Tang Nano 20K and its GOWIN GW2AR-18 FPGA as the main platform, with perhaps a few excursions into other FPGAs.

History and Motivation

Once upon a time, if you wanted to have a custom IC, you went with a wheelbarrow full of money to a semiconductor company. However, some smart person at a semiconductor fab eventually realized they could make a chip with a lot of uncommitted blocks on it and then, for a custom chip, only design the wiring that connected them together. This still required a wheelbarrow full of money, but it was a smaller wheelbarrow.

Then one day, someone realized they could do the same thing but make the electrical connections between the blocks configurable. Maybe have fuses you can blow, or use EEPROM or RAM cells to remember which blocks are connected to which. It is complicated, sure, but then you can make many of these chips and sell them to people who could, in theory, make their own custom chips without your help.

When do you need an FPGA? A classic classroom exercise for an FPGA, for example, is a traffic light because it shows off how to do state machines, which are important for some kinds of FPGA designs. But other than as a learning example, why would you do this? Even a simple 8-bit CPU can handle a traffic light.

Suppose instead that you have hundreds of digital sensors on a rocket, and any one of them must raise an alarm within a few microseconds. A processor has to sample inputs in groups, service interrupts, or rely on extra hardware. An FPGA can simply implement the equivalent of one enormous OR gate. It watches every input continuously, and unrelated logic elsewhere in the FPGA does not steal execution time from it. Can you do it with a microcontroller? Probably, but not easily. For some classes of problems, an FPGA is the better answer.

Of course, you can also build a CPU on your FPGA and some FPGAs have CPUs in the same package. This is often a sweet spot because then things that are easy to do in software, you do in software. Things that are easier to do in hardware, you do in the FPGA.

The 20K Solution

How big of an FPGA do you need? It is hard to compare FPGAs because the blocks are different. Think of PC memory. You can talk about 1 GB vs 4 GB, but the full picture depends on the memory speed, bandwidth, and timing. FPGAs are the same way. Two FPGAs may have a similar number of building blocks, but one may have much more powerful building blocks than the other. However, the 20K board’s FPGA is quite capable, even if it is hard to compare it directly to other FPGAs from other vendors.

It uses GOWIN’s GW2AR-LV18QN88C8/I7, which contains 20,736 four-input look-up tables (LUT4s) and 15,552 flip-flops. The exact maximum clock rate depends heavily on the logic and routing, so there is no single useful “FPGA clock speed,” but the device is quite comfortable in the low hundreds of MHz for appropriately pipelined designs.

Where the chip shines is memory. Most FPGAs have distributed RAM, which eats up resources on the chip, and block RAM, which are dedicated RAM cells. The GOWIN has both of these options. It also has a separate 8 Mbyte, 32-bit SDR SDRAM right in the package. Of course, you have to eat up some FPGA resources to talk to it, but not as much as you might think, and certainly not as much as it would take to place an 8 Mbyte RAM in the FPGA fabric.

There are other special-purpose blocks, like 48 18×18 multipliers, you can configure in a few ways. The Nano 20K board also has a programmable clock generator, QSPI flash (64 Mbits), and connections for audio, LCD, microSD, and HDMI. A microcontroller lets you program the clock generator, flash the chip via JTAG, or provide a virtual serial port back to the host.

Simply FPGA

The basic workflow for an FPGA is:

1. Define the logic you want using an HDL (we will use Verilog, but VHDL is also common).

2. Define a “top block” that connects to the outside world.

3. Set up constraints that mainly tell the device which pins connect to which ports on the top block.

4. Run a synthesis process that converts your HDL to primitives.

5. Run a place and route process that assigns primitives to specific chip resources and produces a “bitstream.”

6. Download the resulting bitstream to the FPGA’s RAM or to the configuration flash (which makes it survive a power cycle).

You can do all of these things with GOWIN’s IDE or open-source tools that we’ll talk about later.

Inside

You might think you just want to draw schematics. Some FPGA tools allow that, but the GOWIN IDE does not. It can, however, draw schematics to show you what you asked for. Turns out, if you are doing something large, schematics are a pain. Why draw out a 7-segment decoder when you can just describe it using a software-like language?

Inside the chip, there aren’t really uncommitted gates per se. There are lookup tables. So if you specify that Z=A|(BC)|(A^D) it doesn’t really go find an AND gate, an OR gate and an XOR gate. It just develops a table and programs that table to produce that function. That’s why LUTs — Look Up Tables — are important.

Simplicity

So let’s look at the simplest possible thing you might do.

module top(output led1, output led, input sw0, input sw1);
assign led=~(sw0|sw1);
assign led1=~(sw0&sw1);
endmodule

We’ll talk more about Verilog later, but for now note that nothing “executes” these lines one at a time. This literally tells the FPGA design software to connect the two switches to a logic gate and then invert the result and connect it to the LED. At run time, this isn’t code, it is wiring.

The LEDs on the board are active low, so this just takes two inputs (presumably switches), computes the OR and AND, and outputs them to the LEDs. The switches on the 20K are pulled down so there’s no need for a pull up or pull down in the FPGA. The output is a logic one when you press the swtich.

How do you connect the devices to your “code?” You need a constraint file, and that’s easiest done through the “Floorplan” editor (see the figure). It builds a file which you can change if you prefer to work with a file:

IO_LOC "led" 15;
IO_PORT "led" IO_TYPE=LVCMOS33 PULL_MODE=UP DRIVE=8 BANK_VCCIO=3.3;
IO_LOC "led1" 16;
IO_PORT "led1" IO_TYPE=LVCMOS33 PULL_MODE=UP DRIVE=8 BANK_VCCIO=3.3;
IO_LOC "sw1" 88;
IO_PORT "sw1" IO_TYPE=LVCMOS33 PULL_MODE=NONE BANK_VCCIO=3.3;
IO_LOC "sw0" 87;
IO_PORT "sw0" IO_TYPE=LVCMOS33 PULL_MODE=NONE BANK_VCCIO=3.3;

That’s it. The schematics come in two flavors. One shows what you’d expect (see the figure). The other shows opaque LUT boxes.

The video below walks through the design from starting the project to watching it work on a breadboard.

Installing

I’m glossing over a lot of details. While the GOWIN software runs on Linux, I had to do a lot to get it to work. Since I wasn’t on a supported distro, I had to work out the dependencies myself, but that wasn’t too hard.

The problem was when I tried to execute the program. The first problem was with libfreetype.so.6. I removed the private copy and linked it to my system copy. That got me a little further. However, to get it all running, I needed to set (or add) GOWIN’s lib directory to LD_LIBRARY_PATH to prevent it from picking up system libraries.

The programmer also needed a little work to not run afoul of other drivers on the system. The specific cure there for me was to build the file: /etc/udev/rules.d/99-tang-nano-20k.rules with these lines:

SUBSYSTEM=="usb", DRIVER=="ftdi_sio", ATTRS{idVendor}=="0403", ATTRS{idProduct}=="6010", ATTR{bInterfaceNumber}=="00", RUN+="/bin/sh -c 'echo $kernel > /sys/bus/usb/drivers/ftdi_sio/unbind'"

Of course, my fixes may or may not help your problems. If you are lucky, you won’t need any fixes at all. The open-source tools are perhaps a little easier to deal with, and we’ll look at those too. However, some of the GOWIN tools are nice, and you can check an option to have it always use your choice of external editors, which is a nice feature.

Next Time

If you need a refresher on digital logic, the original FPGA Bootcamp 0 lets you learn about logic design using a free tool in your browser. No FPGA required at all. Meanwhile, we’ll be looking later at open-source tools, IP, synchronous logic, and using the built-in logic analyzer, among other things. If you want to talk about this series, drop by our Discord server.

When you finish up your next FPGA project, don’t forget to send us a tip. You might see your project on Hackaday.

TL;DR: A malicious Google Ads campaign used fake security warnings to convince Windows and Mac users that their computers had been locked, then pushed them to call fraudulent technical-support numbers. The ads appeared on legitimate, high-traffic sites, including maps, weather, real estate, sports and document-hosting pages. After a user clicked, the ad opened a page that took over the browser and displayed warnings that made it look as though the device had been infected or disabled.

The goal was not to install malware or lock the computer. It was to create enough confusion to get people to call the number on the screen. Callers were then pressured to pay for bogus support, hand over personal information or allow someone to remotely access their device.

Netskope, the security firm that tracked the campaign, said it saw users at 619 customer organizations click on the ads between August 31 and September 14. Its security tools blocked the pages before those users could be scammed. The company said its data represents only a limited view of internet traffic, suggesting the campaign likely reached many more people.

About 62% of the affected organizations were in the United States. Japan and Australia were the next most affected countries. Netskope identified more than 250 Google Ads campaign IDs tied to the operation and found the ads on at least 284 legitimate publisher sites.

What made the scheme effective was the way it behaved in the browser. The page opened in full-screen mode, hid the browser’s address bar and cursor, and interfered with common keyboard shortcuts. The browser also slowed down and played sounds, making it seem as though the entire system had become unstable.

Netskope said the campaign turned a routine ad click into what appears to be a browser or system failure. The fraudulent page occupies the full screen, hides the cursor, blocks common exit commands and slows browser performance to make the device seem unusable. Although the computer is not actually locked, the behavior is designed to alarm users and push them to call the number displayed in the warning.

The fake warnings were adapted for Windows and macOS. The page also waited until it detected mouse movement before displaying the warning. Its code was encrypted and only decrypted in browser memory shortly before the warning was shown. Those choices may have made the campaign harder for endpoint security tools and ad-scanning systems to spot.

The warning pages also made it harder to simply close the browser. Pressing Escape or other keys often had no immediate effect, and attempts to close the page could cause the warning to reload. That behavior is central to the scam: the victim is meant to believe that normal controls no longer work.

Google said it was investigating the campaign. The company did not say why its systems failed to block the ads or whether the entire operation had been removed from its advertising platform.

“We have zero tolerance for scams,” Google said in a statement. “We’re actively investigating the campaigns in this report and will take action against accounts that violate our policies.”

Google has said it blocked more than 99% of policy-violating ads before they were served last year. The Netskope report shows that some campaigns can still get through by using browser tricks rather than a traditional malicious download.

It goes without saying that users who encounter one of the fake lock screens should not call the displayed number. The computer has not been locked, and in many cases, holding the Escape key for several seconds will exit full-screen mode and restore normal controls. Windows users can also simply open Task Manager, while those using Macs can use the Force Quit menu.

Once the browser is closed, it can be reopened without restoring the previous session.

Boston-based voice intelligence startup Modulate has raised $25 in new funding for its platform that uses an array of small models to offer enterprises transcription, emotional analysis, deepfake and AI music detection, and policy enforcement for voice agents in regulated industries.

The funding follows a popular trend among investors in the growing voice AI industry: backing companies that are trying to make AI voices sound more human.

It also rivals other companies trying to detect the intent behind human conversation by analyzing it, and those trying to protect people and companies from deepfake calls, as it is now easy to clone voices.

Modulate’s new funding was led by Future Ventures with participation from Hyperplane and Lakestar. Data from PitchBook indicated that the startup had raised $41 million in funding at a $170 million valuation prior to this round.

The startup was founded in 2017 by Mike Pappas and Carter Huffman, who met as MIT physics undergrads. In its early days, the company focused on providing voice modulation for gaming. But later, it started to concentrate on a voice-based moderation tool.

With the onset of voice AI models, the company is now concentrating on detecting different sorts of AI audio generation and analyzing intent behind a person’s words.

“Our insight into the voice AI space is that a lot of folks are doing transcription, but there’s not really any capability out there that gets the full nuance and full understanding of a conversation, which is so important when you’re talking to another human being,” Huffman said on a call with TechCrunch.

The company today runs more than 100 models that are largely categorized into two sections: Signal extraction models to understand vocal emotion, tone, language, and synthetic voice determination; and Analysis/detection models that look at intent, like what the customer is trying to say, whether the caller is violating rules, or whether they are trying to scam the receiver.

Huffman said that because it runs smaller models, the company doesn’t need specialized hardware and a ton of compute, which could be crucial when token bills go up. Plus, it’s easier for the company to train models with newer capacities, add them to the lot, and have an orchestrator call them when needed.

Modulate has a varied customer base, but it specializes in deepfake detection and alerting organizations like call centers to a possible scam. It also monitors how AI agents respond to customers to assess the quality of calls, along with making sure that AI follows compliance rules in regulatory areas. Because of these products, Module often sits beside the voice stack being used by a company just to analyze calls.

As more enterprises adopt AI-powered customer service, it is becoming important for them to know why a customer call was a success or a failure. In that case, gauging customers’ intent and response becomes critical beyond basic analysis. Huffman said Modulate can give granular data to enterprises around that.

“I think when companies think of emotion analysis, they think if the customer was neutral or positive, the call was a success, and if the customer was negative, the call was a failure. But actually, many times people will be polite even to, like, AI agents or bots. Right. And they won’t come across as angry, but they’ll be very dissatisfied,” he said.

The company said that its tech is also being used to monitor cyberattacks through voice calls.

The startup currently has 40-45 employees and aims to add 10 more people in the coming months to bolster model building. Modulate is currently working on increasing its on-premises and on-device deployment capabilities for increased privacy.

It’s 6 a.m. on a cold January morning in 2002 in New Delhi. It’s still dark outside, and I’m in the kitchen preparing breakfast, packing lunches, and getting my two children ready to catch the school bus when, for the third time in a week, the power goes out. No lights, no mixer to finish my daughter’s puttu—her favorite rice dish—no kettle, no toaster. The bathroom is dark, and the kids are upset.

It will probably be hours before the power comes back on, so I grab a flashlight and light the candles that are set up around the house for these occasions. We’re behind schedule now. We pack the food we have, bundle up as the house turns chilly, and head outside, leaving a mess in the kitchen. We make our way to the bus stop in the dark—the streetlights are out, too—only to discover my daughter has missed her ride. Again. I’ll be late for work at Jamia Millia Islamia, a university where I am a professor of electrical engineering and teach power systems and smart grids. I just hope the power is on there.

This was a common scene for my family and all of Delhi in the early 2000s. Power outages happened almost daily and lasted hours. When the power was on, the quality was so poor that it would dim lights, flicker screens, and wreak havoc on appliances. Customer service at the power utilities essentially didn’t exist.

A child walks in July 2007 past a store in New Delhi specializing in reconditioned generators. The fear of power cuts during summer heat spurs demand for these generators so that residents can produce their own power.Nicholas Bradley/AFP/Getty Images

These problems had been getting worse through the 1980s and 1990s. The cause: an aging distribution grid bereft of crucial technologies, and electricity providers with little accountability. The situation became so bad that the city was losing more than half of its power through obsolete equipment and theft. These staggering losses meant that utilities got paid for only a fraction of the electricity they were trying to deliver. And the lack of funds prevented them from investing in better grid infrastructure.

But over the last quarter century, a remarkable effort by the government and the city’s distribution utilities has turned Delhi’s grid into a reliable, modern system. Power losses have shrunk from over 50 percent in 2002 to 5 to 6 percent in 2026—on par with France and Belgium, and better than Greece and Serbia. Delhi’s grid reliability index, a measure of how often electricity can be counted on, stood at around 70 percent in 2002 and has now topped 99.9 percent.

The bustling Main Bazar in the Paharganj neighborhood of Delhi increasingly uses more nighttime electricity, but reductions in electricity loss help counter demand. iStock

With reliable power, businesses across the city have blossomed. The streetlights are bright. The number of electric vehicles, including city buses, is growing daily. Quality of life has improved. Today, my family is comfortable year-round in our home despite Delhi’s scorching summers and cold winters. The chaos of losing power no longer hinders me from getting to work. The city still has problems—pollution, overcrowding, noise—but thankfully, reliable power is no longer among them.

The transformation of Delhi’s grid can serve as a model for other cities that suffer from decrepit power infrastructure. Regions of Albania, Argentina, Bangladesh, Brazil, Estonia, India, Kenya, Pakistan, Sri Lanka, Uganda, and Venezuela are reeling from heavy losses in their distribution grids. Their problems look like Delhi’s 25 years ago. I believe it’s possible to improve electricity in these places by adapting the changes Delhi made. Here’s an inside look at how the city accomplished it.

Delhi’s Power Grid and Energy Mix

The city of Delhi hosts the capital of the Republic of India, and sits along the Yamuna River in the northern part of the country. It’s home to about 23 million people and is one of the most densely populated areas in the world. Delhi’s grid includes thousands of kilometers of power lines, and peak electricity demand reached an all-time high this year of 8,748 megawatts. The city currently buys 76 percent of its power from central generating companies and private players from neighboring states. Energy generation within the city is restricted to natural gas and renewable sources. Nearly 48.5 percent of the city’s power comes from coal, about 26.5 percent from natural gas, and the rest from carbon-free sources, led by hydropower at 15.6 percent.

Tata Power replaced about 5 kilometers of overhead lines with underground cables, which reduced electricity loss and improved the aesthetics of Delhi’s streets, such as the Janta Flats in the Shalimar Bagh neighborhood.Tata Power-DDL

By the early 2000s, Delhi’s nearly 100-year-old power distribution system was in serious disrepair. Everything was old—lines, transformers, circuit breakers, switches. New grid technologies were needed to keep up with new kinds of electricity loads, but there was little money to upgrade components.

The shabby state of the grid caused many problems, most notably high electricity losses, where electricity vanishes primarily as heat. The cause of the losses was a classic electrical problem: too much current flowing through a network that wasn’t designed to carry it efficiently.

To understand the problem, it helps to understand how modern power grids work. Typically, they include generation, transmission, and distribution. After power is generated, transformers convert the electricity to high voltage levels—typically 132, 220, 400, or 765 kilovolts in India. Transmission lines then carry the power over long distances to receiving substations that are closer to where customers need electricity. Transformers then step down the voltage (to 66, 33, or 11 kV in India) and distribution lines branch out, carrying the power to customers. The whole grid works primarily on alternating current.

In a healthy grid, the utility will take compensatory measures to lower the current and maintain the voltage all the way to the ends of the lines. But in Delhi, this wasn’t happening. The result was a vicious cycle. Reactive loads increased the current, the higher current increased energy losses and lowered the voltages, lower voltages forced devices to draw more current and further increased the losses.

In some parts of Delhi, the effect was so severe that residents took matters into their own hands. A colleague of mine who lived in a different part of the city constantly experienced voltage that was too low for her appliances to operate reliably, so she had to install her own voltage stabilizer. At my home, we bought an inverter and battery system to keep a fan and a few lights running during the many outages.

Electricity Loss and Theft in Delhi

The losses in Delhi weren’t caused solely by technical problems. Theft of electricity was rampant, by both the powerful and the powerless (in both senses of the word). Businesses, residential customers, and utility employees with vested interests would siphon electricity from the grid. It was easy to illegally hook into a streetlight or a distribution line running close to one’s house or factory. Utilities didn’t have the resources to identify theft or penalize offenders. Even if they could, the courts were already overburdened, and an electricity regulatory commission that could push for reforms had not yet fully formed.

Updated meters have made billing easier and more accurate. Tata Power-DDL

Making matters worse, the utilities and their employees were rarely held accountable for their actions, and so corruption plagued the system. Junior engineers and line workers, many of them lacking appropriate technical skills, were tasked with handling nearly every issue, including outages, flickering, and bill payment. This was too much authority in the hands of people with too little training.

On top of that, customers didn’t pay their bills. Meters were old, frequently faulty, and easily tampered with. Utility employees would take a meter reading by visiting the customer’s property, noting the reading in a book, entering it in a ledger or on a computer back at the office, and converting it into an electricity bill that would get dropped off at the customer’s property. This process left a lot of room for incorrect billing.

To pay a bill, customers had to stand in long queues at the utility offices, which had limited business hours. Not wanting to take off a half day of work for this, many customers simply didn’t pay. And there was no penalty for not paying—there were no regulations allowing the utilities to cut off a customer’s power. (I paid my bill by having a family member stand in line for me.)

The combined commercial and technical losses left Delhi’s utilities collecting payment for less than half of the electricity they were supplying in the early 2000s.

India’s Electricity Act and Power Reforms

Such problems weren’t unique to Delhi. On average in 2002, state utilities across India experienced electricity losses of nearly 37 percent. My country desperately needed systemic reforms, but authority over electricity was split between the central and state governments so any decision-making was fractured. States managed most of the generation, as well as transmission and distribution, while the central government oversaw generation that supplied multiple states, such as hydropower, fossil fuel plants, and nuclear plants. The central government could push reforms, but the states determined whether those reforms would succeed. Making matters worse, most states put a single organization in charge of generation, transmission, and distribution, giving that entity too much control and reducing transparency and competition.

A team of technicians with BSES Rajdhani Power maintains an insulator string on a large power transformer in 2011. BSES Rajdhani Power

In 2001, India’s central government began writing some historic legislation that became the landmark Electricity Act, 2003. Among the grand reforms aimed at transforming the country’s power industry, it unbundled state oversight of grid networks, creating separate entities for generation, transmission, and distribution. It also opened up the power sector to privatization. It allowed large electricity customers to bypass local distribution companies and purchase electricity from competitors or build their own power plants. It created a central regulatory agency responsible for determining interstate tariffs and promoting market competition in the power sector. And it created mechanisms for prosecuting electricity theft.

Hundreds of capacitor banks have been installed in Delhi to supply reactive power at strategic locations and help stabilize voltage.Tata Power-DDL

In 2002, Delhi was already taking drastic action to fix its grid. The organization overseeing Delhi’s distribution, the Delhi Vidyut Board, was broken up and two private companies—BSES (now Reliance Infrastructure), and Tata Power—took over distribution. They faced a Herculean task. Tata Power, serving the northern half of Delhi, would have to tackle a combined commercial and technical electricity loss of 53.5 percent. BSES, whose territory was split between two subsidiaries, was facing 51.5 percent losses in South Delhi and 63.1 percent losses in East Delhi.

“The company inherited a deteriorated and overloaded network, massive power theft, weak billing and collection systems, inaccurate consumer records, and an aging, largely untrained workforce,” Dwijadas Basak, CEO of Tata Power, told me. There were over 100,000 unresolved billing complaints, 20,000 pending connection applications, and frequent supply failures, which had severely eroded consumer trust, he added. Both Tata and BSES devised sweeping reforms and human resource development initiatives. The companies followed their own paths over the years, but ultimately implemented similar changes, with similar results.

Delhi’s Electricity System Overhaul

Fixing Delhi’s grid was a journey that involved all stakeholders, including customers, city authorities, and utility employees at all levels. The utilities revamped their organizational structures, diminishing the power of junior staff and creating separate teams to focus on specific tasks. Long-term employees of the erstwhile Delhi Vidyut Board received training from the up-and-comers at the new companies.

On the technical side, both companies installed digital control systems that let them monitor and operate the grid from a central location. Known as SCADA, or supervisory control and data acquisition, the systems offered a bird’s-eye view of the infrastructure, including the status of equipment, voltage, current, power flow, and switch positions, with updates in seconds. This helped the companies identify areas of high loss and theft and make faster decisions based on accurate information.

The SCADA (supervisory control and data acquisition) system at Balaji Estate in Delhi’s Kalkaji neighborhood serves as the nerve center of BSES Rajdhani Power’s distribution network in South and West Delhi. It enables real-time visibility, remote control of grid operations, fault identification and isolation, and load management. BSES Rajdhani Power

The utilities also replaced aging transformers and circuit breakers and created extensive maintenance plans for equipment. In 2002, 11 percent of the transformers in the region were failing at any given time. That rate is less than 1 percent today, according to Tata. Crucially, the companies installed hundreds of capacitor banks, including some mobile ones, to supply reactive power at strategic locations. This improvement reduced the total current flowing in the distribution lines and helped stabilize the voltage. They also installed voltage regulators at points in the system where voltage tends to drop.

To reduce theft, the companies replaced bare distribution wires with insulated lines—a single cable for three phases—which made it harder to tap into the lines. The cables also reduced outages because they’re better at preventing ground faults, which can occur when, say, a tree branch falls on the line.

Workers received better sensors and tools to do their jobs safely and accurately. For instance, they were given helmet-mounted voltage sensors, which are safer than handheld ones, and thermal scanning tools to detect hidden defects in the insulation of high-voltage equipment that could otherwise have led to catastrophic failures.

To reduce inaccurate billing and meter tampering, the companies replaced the old electromechanical meters with digital ones that are read with handheld devices. In some locations, radio-frequency-based group metering systems were installed by Tata to consolidate multiple customers’ meters into one. The data is then wirelessly transmitted to a central database, eliminating the need for individual meter readings. The companies are now trying smart meters, which give consumers more control over their electricity bills and give utilities remote control of some equipment (with the customer’s consent).

To encourage people to pay their bills, the utilities installed kiosks that are available 24 hours a day, and they created a web-based payment system and mobile app. Incentives for early bill payment and community-engagement programs also helped. Assistance from Delhi’s law enforcement considerably reduced electricity theft.

Tata Power hired women living in the 223 slums it serves in the northern parts of the city to knock on neighbors’ doors and remind them to pay their power bills. These payment collectors [left and center], known as abhas, were photographed while speaking with a customer [right] in the Sanjay Basti area of New Delhi in 2017. Prashanth Vishwanathan/Bloomberg/Getty Images

In areas where theft was particularly rampant and losses were as high as 83 percent, according to Tata, the companies took a different strategy. These pockets of Delhi were predominantly occupied by low-income families. Tata Power, and later BSES, worked to improve the water supply for these residents and provide educational opportunities, such as instruction in reading and writing in Hindi as well as financial literacy. These efforts focused on the women, who were at home more, and paid them to collect electricity payments from their neighbors. Bill payment rates from these areas are now on par with those of other parts of Delhi.

In recent years, some customers have been installing rooftop solar panels to take advantage of subsidies and incentives. This trend can reduce electricity losses further because the energy generated at the customer end reduces current in the distribution lines. Customers are also installing more LED lights and energy-efficient appliances, reducing the load in the system.

BSES is using AI to help detect theft. The algorithms analyze consumption patterns in pockets where losses are higher than they should be. The company is also using AI to forecast demand, fine-tune operational efficiency, and provide chatbots for customers.

Quality of Life Improves in Delhi

Life in Delhi is better than it was 25 years ago. I’m not worried that the power may go out and force me to reschedule my activities. My uninterrupted Wi-Fi gives me peace of mind, and my heating and cooling systems keep me and my family comfortable. I rarely need to use our old inverter and battery.

The sharp rise of e-rickshaws in Delhi has increased demand on the power grid. Sajjad Hussain/AFP/Getty Images

The number of businesses in Delhi has increased substantially, in part because of the access to quality power. People can confidently buy products that depend on electricity. In fact, the city’s peak electricity demand has tripled since 2002 due to the increase in population, commercial activity, and use of electrical gadgets.

And then there’s the benefits to the planet. One unit of electricity that isn’t frittered away is one less unit that must be generated, not to mention the reductions in carbon emissions.

Still, there’s work to do. Some areas of Delhi continue to have high losses, driven partly by the illegal charging of e-rickshaws. Elsewhere in India, the states of Himachal Pradesh, Madhya Pradesh, Maharashtra, and Telangana still experience losses of about 17 to 23 percent despite the sweeping Electricity Act, 2003. There are many reasons for the ongoing losses: long distribution lines to remote villages, less digitization, and inefficiencies in billing and collection of payments.

These regions, and others around the world, can learn from Delhi’s grid comeback. Recently, power losses have increased substantially in countries such as Argentina, Greece, Jamaica, and Morocco, according to the World Bank, and some of the causes are similar to those that Delhi faced back in 2002.

Meanwhile, Australia, most countries in North America and Europe, and a few countries in Asia and Africa experience low electricity losses as they invest regularly in their distribution infrastructure and the ethical enforcement of rules. In China, for example, losses have gradually been cut in half, from 7.1 to 3.4 percent. In Latvia, losses plummeted from 25 to 5.8 percent.

What’s important is a comprehensive approach. Technologies like smart metering, AI, and analytics certainly help, but equally important is that people in the field are trained and take responsibility for their jobs, and that laws are enforced and payments collected.

“Sustainable loss reduction cannot happen through technology alone,” Abhishek Ranjan, CEO of BSES Rajdhani Power told me. “Technology is an important enabler, but long-term success comes from combining it with disciplined execution, operational accountability, and strong consumer engagement.”