This is the latest arrest in a months-long investigation conducted by Taiwanese authorities over suspicion of illegal exports.

A Taiwanese national, reported to be an Nvidia employee, was arrested in the country on Tuesday (28 July) on suspicion of illegally exporting high-end AI servers manufactured by Supermicro.

According to a translated statement from Taiwan’s Keelung district prosecutor’s office, the suspect, named Chang, is accused of crimes including making false entries in business documents.

Investigation into his activity began on 24 July, and authorities conducted searches of his residence and business. Chang was later brought in for questioning.

Authorities detained Chang after determining him to be a flight risk. Concerns were also raised around the suspect potentially destroying evidence or colluding with accomplices.

The prosecutor’s office did not mention Nvidia in its statement, although multiple news publications have reported that the Supermicro AI servers in question were equipped with Nvidia chips, which are subject to US export controls and cannot be transported into China.

Nvidia did not confirm whether the detainee was its employee. “We primarily sell our products to well-known partners, including ​OEMs (original equipment manufacturers), who help us ensure that all sales comply with US export control rules,” the company said in a statement to Reuters.

This is the latest arrest in a months-long investigation conducted by Taiwanese authorities over suspicion of illegal exports, and came amid a third round of searches. Earlier in July, Supermicro said that two of its workers at its Taiwan unit were arrested as part of the investigation into the alleged illegal exports, while a previous round saw three people being detained.

In March, the US government charged three people associated with Supermicro, including co-founder Yih-Shyan Liaw, over allegations of helping smuggle at least $2.5bn of AI technology into China.

Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.

Given the meteoric rise in BGMI players, Krafton India has just announced the BGMI Campus MVP Program. This 90-day campaign will have 150 student ambassadors selected from 17 IITs and 28 NITs. The objective being to build an active BGMI community among the participants by getting them involved in organizing tournaments and events. KRAFTON also plans to combine online gameplay with offline meet-ups to keep students engaged throughout the program.

What Is BGMI Campus MVP?

Campus MVP BGMI is an initiative by KRAFTON India that aims to create gaming communities among college students. Under the 90-day campaign, certain students would be appointed as BGMI ambassadors at their respective campuses. These students would promote gaming among other students by organizing tournaments, creating content, and conducting community activities. This campaign attracted over 16,000 applications from students from 5,000 different institutions. KRAFTON selected 150 students for the first batch, while an additional 100 students were selected as reserve candidates.

The chosen student ambassadors will conduct BGMI sessions within their respective colleges during the 90-day program period. They will arrange classic matches, scrims, and tournaments, generate gaming content for college students, and also hold offline meet-ups. The ambassadors will conduct inter-department and inter-college competitions before participating in national college gaming competitions.

The first phase of the Campus MVP program is already in progress. After completing onboarding and orientation, student ambassadors have begun organizing BGMI matches, creating gaming content, and running campus events. The initiative follows a three-phase format over 90 days.

KRAFTON’s Vision Behind the Initiative

According to KRAFTON, BGMI is no longer just about playing matches. KRAFTON claims that the game has turned into a platform where players can not only challenge each other but also create content and network with people. In this way, with the Campus MVP program, KRAFTON hopes that students will be able to network through games as well as take up leadership positions at college. This program is one of the many ways in which the company aims to strengthen ties with young gamers in India. KRAFTON also plans to use the learnings from the first batch to expand the program to more institutions in the future.

Apple just gave you a new way to own an iPhone, iPad, Mac, or Apple Watch without buying it. The newly launched Apple Upgrade program lets you lease Apple hardware through Klarna instead of paying full price upfront, and I’ll admit that the monthly numbers genuinely look tempting.

I went through Apple’s own published rates device by device to figure out where this actually saves you money, and where it’s just a slower way to pay full price for something you’ll never own.

Who qualifies for Apple Upgrade?

Apple Upgrade relies on Klarna as the lease provider. Applying for a device only triggers a soft credit check. To apply, you need to be 18 or older, a US resident, verified by SMS, and carrying an eligible credit/debit card, along with your Social Security number.

As part of the program, iPhones and Apple Watches are leased over 12 or 24 months, while iPads and Macs stretch to 24 or 36 months.

Eligibility covers most current models, except the iPhone 16, iPhone 16 Plus, Apple Watch SE, MacBook Neo, Mac mini, the base iPad, and Studio Display.

iPhones also require an active AT&T, T-Mobile, or Verizon plan, though the phone itself ships unlocked. AppleCare+ isn’t bundled anymore; it’s billed separately.

What happens when your lease ends?

Once the lease ends, you actually have three options based on what you want to do with the device: return and upgrade, pay to own, or walk away empty-handed. 

Apple gives you up to six months after your term ends to decide. If you do nothing, Klarna automatically charges that purchase-option fee on your behalf.

iPhone: Does leasing actually save you money?

Let’s consider the iPhone 17e at $599 on a lease-and-return basis. If you lease it for 24 months and pay $17.99 a month, you’ll hand over roughly $432 total before sending the phone back. Basically, you never owned anything; you just rented it for two years.

If you decide to keep the phone instead, you’ll pay the difference, roughly $165 to $170 more on the 17e. Add that to your lease payments, and the total lands right around $599, the exact price you’d have paid on day one had you purchased the iPhone outright.

To me, that’s where Apple Upgrade’s real appeal lies. It gives you the flexibility to spread out the cost, keep your cash available for other priorities, and still have the option to own the device later. However, I think the program makes far more sense for expensive Pro models than entry-level iPhones, where the upfront cost is easier to justify.

iPad: Worth the monthly commitment?

Given that buyers tend to keep their iPads longer than iPhones, Apple provides longer tenure options to lease them: 24 or 36 months. If you’re a student or an entry-level professional who simply wants an iPad for studies or getting done with the day-to-day digital chores, going for the lease makes good sense.

For one, you’ll save the $599 (for the iPad mini) or the $749 (for the iPad Air) you’ll spend upfront, pay $11.99 or $15.99 per month over 36 months, and then, depending on whether you need those devices, can either pay the difference or simply upgrade to the latest one without paying any additional money.

Again, there are no discounts here: just ease of usage.

Mac: A better deal or just a bigger bill?

If you don’t have that kind of money upfront but still want to upgrade from an old Intel-based MacBook or even an M1- or M2-powered MacBook Air, the Apple Upgrade program breaks that upfront cost into monthly installments starting at just $24.99 over 36 months.

However, there’s another side to this. The M5 family of chips offers enough performance headroom to stay relevant for the next four to five years, which is why I expect these MacBooks to hold their resale value exceptionally well. 

That’s why I’d stretch the payments over the full 36 months and then pay the remaining amount to keep the MacBook. Unlike an iPhone, a MacBook is something that I’d comfortably hold on to for four or five years, choosing long-term ownership over upgrading.  

Apple Watch: The one category where leasing might win

I’d personally say that it makes more sense to buy the Apple Watch instead of leasing it.

First, the Apple Watch upgrade trajectory hasn’t really been on an upward path lately. For most users, the Series 10 is going to serve the same purpose as the Series 11. With a new release around the corner, you can generally find the outgoing models at a decent discount.

Take me as an example: I’ve been using the Watch Series 8 for three years now, and I don’t really feel like upgrading. The upfront cost isn’t as much as a Pro iPhone or MacBook Air either.

from the 30-days-of-wtf dept

For a long time, cell site simulators, a.k.a. “Stingrays” made headlines on nearly a daily basis. Then they just kind of fell off the map.

L3Harris — the manufacturer of cell site simulators that commanded enough market share to see its flagship product become the victim of genericide — saw the writing on the wall and exited the market. Part of this was due to cell tech advances that made it more difficult to obtain the information these faux cell towers were meant to collect. Part of that was also Supreme Court precedent that made the tech inherently less popular with US law enforcement.

Riley raised questions about cell site simulator use by requiring warrants for cell phone searches. And spoofing a cell tower was definitely a search, as the devices forced every cell phone in the area to connect to the Stingray and cough up identifying info about the device. Carpenter arrived a few years later and made it clear long-term tracking via cell site location was no longer something covered by the Third Party Doctrine.

But the biggest contributor to the decline in Stingray device usage were the warrant requirements instituted by both federal and local law enforcement agencies. What used to be a Wild West free-for-all was now something that required judicial approval. Apparently, a lot of cops decided this tech they once claimed was so useful to investigations it couldn’t be discussed in open court was useless now that it was subject to oversight.

We’re seeing a bit of an anomaly here. Not only is the use of cell site simulators being discussed in open court, but the federal officers have been denied their unusual — and outrageous — request to basically go wardriving for a month in Akron, Ohio. Here’s This Week in Security with more details:

A U.S. magistrate judge last month denied to issue a search warrant allowing federal agents to snoop on the phones of “thousands of uninvolved, unsuspecting individuals” across Ohio in an effort to identify a suspected criminal’s device, a rare rebuke by a court blocking the use of a cell-site simulator.

In the ruling, the judge said the federal government wanted to deploy a cell-site simulator that would have allowed “access to the information of thousands of individuals in the Akron, Ohio area,” but refused the warrant on grounds that it would have allowed federal agents to “gain unbridled discretion to examine the movements of private citizens at all times for thirty days.”

The ruling [PDF] by the magistrate doesn’t name the federal agency seeking the warrant, nor does it give any details about the alleged criminal activity that might help narrow the list down. The rest of the docket remains sealed so it may be weeks, months, or never before we learn anything else about this incident.

Here’s what it does say about the events leading up to this severely deficient warrant:

On June 15, 2026, the government sought approval of a warrant for use of CCSS for up to 30 days, 24 hours a day, to identify the cellular device(s) used by a suspect involved in criminal activity in Akron, Ohio. The affidavit to the warrant application established probable cause to believe that a specific individual is using one or more unknown cellular devices in criminal activity. The affidavit also suggested that there is probable cause to believe that uncovering the identity of the cellular device(s) would unveil more evidence of criminal activity.

The problem is right there in the first sentence. The government appeared to think the only thing it had to do to satisfy the particularity requirements of the Fourth Amendment was to suggest it might limit this roving, 24/7 surveillance to a few areas in Akron.

Moreover, in an attachment to the warrant application, the government described five different locations at which a CCSS could be used to identify the suspect’s cellular device(s) “when the officers to whom it is directed have reason to believe that [the suspect] is present” and “in the
vicinity of” the following locations:

the suspect’s residence;
the suspect’s overnight location;
the suspect’s daytime location; and
two other densely populated locations the suspect frequently visited.

That’s wild. This is basically telling the court the government wants to force thousands of devices to connect to its cell site simulator at multiple locations for a period of 30 days. That the agency said it would “take no further investigative steps” until it had gathered enough info to make sure it had found its preferred suspect is hardly comforting. The fact that it claimed it would delete any irrelevant information (at an unspecified time) following its 30 days of wardriving doesn’t help much either.

The court says this is obviously impermissible under any interpretation of the Fourth Amendment, especially given Supreme Court precedent handed down in recent years. It also cites geofence warrants that have been recently rejected by magistrates for pretty much the same reason: wholesale surveillance of hundreds or thousands of people attempting to present itself as a legitimate search under the Fourth Amendment.

The court reminds the government that the Fourth Amendment says this about warrants:

[N]o Warrants shall issue, but upon probable cause, supported by Oath or affirmation, and particularly describing the place to be searched, and the persons or things to be seized.

The court says this warrant is no better than the geofence warrant rejected by an Illinois federal court back in 2023: searching for a suspect by searching everyone officers encounter isn’t “particular,” and the use of warrant doesn’t make it more acceptable.

There, just as here, the proposed warrant allowed the government access to thousands of identifiers and location data in an undetermined geographic area. Id. at 715–16. The court found the CCSS [canvassing cell site simulator] “akin to a warrant application to search an entire apartment building—or an entire city block—where the government has probable cause only that evidence of a crime will be found in one specific apartment unit, for up to 30 days.” As that court recognized—and as this Court now concludes—the Fourth Amendment’s particularity requirement bars this sort of “rummaging” through the proverbial home.

That much should have been obvious to the investigators seeking to have this warrant approved. Either this was ignored in hopes investigators could slip one by the judge or the government thought this might somehow be more constitutional than a geofence warrant with the same parameters. Either way, the government was wrong.

Despite cops relying on Google location data more than cell tower dumps or Stingray devices these days, it’s clear they’re still relying on tech that has completely fallen out of favor over the past several years. I guess if you’ve already bought it, you may as well use it. Sunk cost meets diminishing returns. Fortunately for the Fourth Amendment, blowing the dust of a cell site simulator hasn’t changed the way courts view these warrants.

Filed Under: akron, cell site location, cell site simulators, ohio, stingray, surveillance

Summer temperatures keep pushing higher across the map, and grids feel the pressure. Air conditioners run nonstop, transformers strain, and outages arrive with little warning. In those hours a reliable backup power station stops being a nice-to-have and becomes the difference between a minor inconvenience and a real problem. The DJI Power 1000, priced at $349 (was $699), steps into that role with a combination of capacity, output, and practical design that matches the needs of actual emergencies.

A 1024 watt-hour Lithium ferrophosphate battery is packed into a unit that weighs only 29 pounds and measures 17.6 by 8.9 by 9.1 inches, making it compact enough to transport from room to room or fit into a vehicle if you ever need to make a rapid departure. Size isn’t everything; the relationship is what really matters. LFP cells have a reputation for delivering over 4,000 charge cycles while keeping more than 70% of their original capacity, equating to nearly a decade of regular use without significant drop down.

Sale

DJI Power 1000 Portable Power Station, 1024Wh LiFePO4 Battery, 2200W (Peak 2600W) AC/140W USB-C Output…

• Power 99% of Household Appliances – With a 2200W max continuous output power, DJI Power 1000 easily handles high-watt appliances like outdoor camping…
• Fully Recharged Fast – DJI Power 1000 can be fully recharged in just 70 minutes using grid power. Charge fast and save time so you get moving on your…
• Safe and Secure – An LFP battery provides up to 4000 cycles and a service life of approximately 10 years. The DJI Battery Management System (BMS…

Continuous AC power reaches 2200 watts with a short-term peak of 4400 watts, which is handled by two normal household outlets. A typical fridge or freezer consumes less power than that continuous rating, therefore the station should be able to keep your food safe for an extended duration. Medical gadgets, fans, phone chargers, routers, and laptop computers all remain operational without incident. Two USB-C connectors supply up to 140 watts each, enough to charge current laptops and phones simultaneously, while two USB-A ports power older devices and accessories.

When the power is restored, even momentarily, recharging occurs quickly, with 80 percent completed in approximately 50 minutes and a full charge in about 70 minutes under ideal conditions. That speed allows your station to recover from periodic outages rather than sitting idle for hours, which is an extra benefit. If you do have sun, the DJI adapter module is an alternative for getting through longer events. The unit’s quiet functioning at 23 dB allows it to sit in a bedroom or living room overnight without adding to the stress level.

In UPS mode, the station ensures a smooth transition. When you put it into a wall outlet and your gadgets consume power from it, a grid breakdown activates the battery power in roughly 0.02 seconds, with no interruption, data loss, or worry. Bypass charging keeps the battery charged while supplying power, ensuring that the system is ready for the next reduction in power.

An intelligent battery management system monitors for overcurrent, overvoltage, and overdischarge concerns. Individual circuits are protected by multiple fuses. The bright color display reveals remaining capacity, input and output levels, and status at a glance, allowing you to monitor things and make educated decisions even in low light conditions.

Spur Intelligence, a cybersecurity startup based in Lake Mary, Florida, has raised a $200 million round led by Insight Partners.

Spur, founded by two former Defense Department engineers in 2017 — five years before ChatGPT’s public launch — was prescient. The startup’s tech helps enterprises distinguish legitimate human users from increasingly well-hidden bot traffic to help identify fake users and threats.

“As sophisticated criminal VPNs, residential proxy networks, and anonymization infrastructure proliferate, organizations are increasingly operating with a critical blind spot: they can see the activity, but not the infrastructure behind it,” Insight’s Thomas Krane said in a written statement.

Detecting malicious traffic has, of course, been a hill corporate security teams have been climbing for eons. But nothing compares to the onslaught facing them today. As of mid-2026, bots are now more active on the internet than humans, Cloudflare reported last month.

“Thought it would be end of 2027, then early 2027, but agentic traffic growing so fast that bots have now passed human traffic online for the first time in the Internet’s history,” Cloudflare founder and CEO Matthew Prince posted on X last month, pointing to his company’s latest traffic report.

Security

The vendor won’t confirm or deny

We now have a better idea of how OpenAI’s models broke out of their cages to attack Hugging Face. The rogue models found zero-day vulnerabilities in JFrog’s universal binary repository manager Artifactory around the time they escaped, according to JFrog CTO Yoav Landman.

While Landman doesn’t outright admit that the JFrog flaws were the zero-days that OpenAI’s models found and exploited, ultimately allowing them to breach the massive model mart, it definitely looks and quacks like a duck – err, frog. 

Landman says OpenAI’s models discovered the Artifactory zero-days during a security evaluation. The AI giant notes the incident occurred while its models were being evaluated on the ExploitGym benchmark.

“During a security evaluation, OpenAI’s models identified previously unknown zero-day vulnerabilities in self-hosted Artifactory installations that could be exploited to gain unintended internet access,” Landman said on Monday.

JFrog Artifactory is a central platform that organizations use to store and distribute all the software artifacts across their supply chains. It supports more than 60 package formats including Docker, Maven, npm, PyPI, Helm, and AI/ML models. 

OpenAI “responsibly and immediately” disclosed the vulnerabilities to JFrog, Landman continued. “Our security team treated the report with the urgency it deserved, as a genuine zero-day unknown to the world, and moved accordingly. We developed, validated, and released a fix for all JFrog customers, self-hosted and cloud alike.”

On Monday, JFrog released the fixed versions, and credited OpenAI researchers for reporting at least eight of the now-patched Artifactory vulnerabilities: CVE-2026-65617, CVE-2026-65925, CVE-2026-65921, CVE-2026-65923, CVE-2026-66018, CVE-2026-66014, CVE-2026-66015, and CVE-2026-65924.

The Register asked JFrog whether at least some of these were abused by OpenAI’s rogue models to access the internet and compromise Hugging Face, but the DevOps firm isn’t talking.

JFrog’s admission comes about a week after OpenAI said two of its models, GPT-5.6 Sol and a second pre-release model, escaped their testing sandbox during a security evaluation designed to test their cyber capabilities. During this test, the models found a way to access the open internet, then broke into Hugging Face and accessed private information and stole some credentials.

“While operating in our sandboxed testing environment, our models spent a substantial amount of inference compute finding a way to obtain open Internet access, in pursuit of solving the evaluation problem,” OpenAI said on July 21.

“To gain access, the models identified and exploited a zero-day vulnerability (which we’ve now responsibly disclosed to the vendor) in the package registry cache proxy.”

The Register also reached out to OpenAI and asked if JFrog is the vendor referenced in its blog, but did not receive any response. 

Luckily, your humble vulture is headed to Vegas in a week for Hacker Summer Camp, so we will be sure to test our betting prowess in the appropriate environment. Without guardrails, of course. ®

Ask ChatGPT for a chapter in the voice of Stephen King and it now says no. It offers instead to write something with “the hallmarks of atmospheric, character-driven horror and small-town dread,” while staying “its own.”

OpenAI has quietly changed how its chatbot handles author imitation, Ars Technica reported after testing it. The model declines to copy a named writer’s “exact style” and redirects to broad craft traits instead.

The interesting part is not that it refuses. It is when the refusal started, and how far it now reaches.

A boundary that moved in two weeks

On 15 July, the research outlet No Latency published a comparative audit of five chatbots across 35 test responses. It found ChatGPT refused to imitate living authors but still complied for dead ones, including recently deceased writers.

By 27 July that gap had closed. Ars found ChatGPT dodging requests for dead authors too, naming Charles Dickens and Ernest Hemingway alongside living writers like J.K. Rowling.

Engadget saw the same thing a day later. Asked to write a mystery in Agatha Christie’s style, the chatbot replied that her “works are still under copyright, so I can’t provide text that closely imitates her distinctive style.”

Christie died in 1976. The mention of copyright, not mortality, is the tell. OpenAI appears to have shifted the line from “living authors” to “anyone still in copyright.”

Why the wording matters in court

The change reads less like a craft decision than a legal one. OpenAI is fighting a stack of copyright suits from authors and publishers who say their books were used to train ChatGPT. One complaint cites the model’s “uncanny ability to generate text similar to” copyrighted work.

US copyright protects a specific expression of an idea, not the looser notion of style. So a chatbot that refuses a writer’s “exact style” while still offering the general “feeling” is standing on the safe side of the line. It withholds the reproducible expression and hands over the unprotected mood.

That subtlety could still matter. Legal scholars note that an imitation can become infringing if it grows “substantially similar” to the original.

“We’ve never had a situation in which this personal style of individual creators could be imitated as well and as inexpensively as we now have with AI,” Robert Brauneis, a law professor at George Washington University, told Bloomberg Law.

Every lab draws the line somewhere different

The No Latency audit found no shared industry rule, which is the strongest sign this is policy rather than capability.

On living-author prompts, ChatGPT and Perplexity refused and redirected. Anthropic’s Claude and Microsoft’s Copilot complied, but attached caveats about originality. Google’s Gemini complied outright, with no visible hedge, even as it faces its own publisher copyright suit.

The models can all do the task. They simply disagree on whether they should. As the study put it, AI companies are drawing product boundaries around creative identity, deciding case by case whether a name is an influence, a teaching example or an impersonation.

OpenAI has been here before with pictures. Its DALL·E 3 image model already declines to render a living artist’s style. Yet its published model spec from December carries no equivalent written rule for text, which is why the new refusals look like a quiet retrofit rather than a stated policy. OpenAI did not respond to Ars Technica’s request for comment.

What it means for the people who used it

For writers who leaned on the feature, the shift is already an irritation. “Now Ms. GPT says she can’t generate content in the style of specific authors,” one user wrote on Reddit. “My prompts were soooo specific and I got exactly what I wanted out of them.”

The Authors Guild sits on the other side. Its best-practice guidance urges writers not to use AI to “copy or mimic the unique styles, voices, or other distinctive attributes” of others, warning of unfair-competition and infringement claims.

The friction is the point. OpenAI is trying to keep a useful writing tool while narrowing the one use that most invites a lawsuit. Its rivals have not made the same call, and its own courtroom exposure keeps growing, from the newspapers suing over training data to the record settlements now setting a price on the practice. Licensing deals like Getty’s pact with OpenAI point at one exit. A disclaimer bolted onto a refusal points at another, cheaper one.

Instacart is posing the provocative question: What if most of the work your engineers do today should, in fact, be done by machines?

At VB Transform 2026, CTO Anirban Kundu argued that dev teams continue to waste their time on draining, repetitive, high-volume work; this should be absorbed by AI agents so that humans can focus on problems that require judgment, intent, and exception handling. 

In fact, in 97% of cases, Instacart’s builders don’t even read code anymore. 

“In the past, the tactical level was the creation of the code,” Kundu said. “In the most tactical level going forward, it’s going to be, ‘How do you navigate around the AI system to give you what you want?’”

AI generating code, performing “pretty serious evals” 

That doesn’t mean humans never look at code; agents handle the bulk of code generation and boilerplate, particularly with newer projects where code is generated or regenerated on a weekly basis. 

“The benefit of that is we don’t care about tech debt anymore,” Kundu said. “Things that are not active just get dropped out and then it gets rebuilt, kind of like how we used to build assembly code or object code.”

So why not 100%? The remaining 3% is in legacy, compliance, and latency-sensitive systems and workflows, or driven by a “boatload of code” that is dead, not active, or half-active. These cases still need careful human attention. 

Instacart is slowly “smoothing those parts out,” however, breaking systems down in an aptly-named project Atoms, then building them back up in a cleaner, more modular form. Kundu’s team started with the “monoliths” and is shifting to remote procedure call (RPC)-driven architectures. 

But evaluation remains one of the overarching challenges. Code reviews aren’t as relevant when AI is generating code — as Kundu noted, “the lines of code are going to be correct, the syntax is going to meet your expectations” — so the goal is to move to an “intent model.” That is, training devs so they can ask different models the right questions from an intent perspective. 

Evals are then performed independently: Roughly 7,000 automatic evaluations run each month, and the system answers 8,000-plus real-time developer queries with about 99.9% accuracy.

Identifying “hiccups” that human intuition might have missed

Dovetailing with this, Instacart has built an agentic site reliability engineering (SRE) system trained on years of the company’s own incidents and root-cause analyses rather than generic failure data. Instead of teaching a model how production outages work in the abstract, the team fed it the specific ways Instacart’s systems have broken over time, along with the ways humans diagnosed and fixed them.

As a result, the company has seen accuracy in detecting and mitigating production issues jump from roughly 60 to more than 90%. 

Kundu pointed to one example with Instacart’s internal tool Blueberry. The AI SRE colleague watches 200-some-odd Slack channels, monitors signals, and looks for patterns across human conversations and alerts. 

In one incident, a database shard backed by an EBS volume that had a “hiccup” for a period of time. The human team did not immediately suspect AWS disk issues and were “obviously scrambling” to figure out why this particular shard misbehaved. 

But about 20 minutes in, Blueberry posted on Slack, pointing to a specific blip and tying it to a feature-flag-like system called “roulette” that had been inadequate. “It’s supposed to be rolling out in this cadence, [but] it had been too much,” Kundu said.

Blueberry figured it out, and the team resolved the incident. “Would have a human been as quick? I think the problem is human intuition would hold us back a little bit,” Kundu said. 

Humans tend to default to patterns we’ve seen before, then resort to debugging; Kundu called this the “first brain-second brain kind of thing.” But Instacart’s agentic SRE is actually “more comprehensive in its ability to look at everything and then be able to decide what does or doesn’t matter.”

Redefining the engineer’s job

Looking ahead, the most tactical work for engineers will be navigating AI systems: Designing and supervising evaluation processes; coordinating multiple simultaneous experiments and features; managing constraints like limited top-of-funnel traffic for testing; figuring out when to escalate; identifying edge cases and where things might break.

Domain expertise is also being rethought in the age of AI. Instead of bottlenecking changes through a single “owner” team that touches the code, Instacart is embedding domain knowledge into definitions and specs that any team can use. 

“We’ve lived in this world where this group or this engineering team is the one that can touch the code and make the modification,” said Kundu. “We’re trying to move into a world where the code becomes completely democratized across groups.”