BUSINESS

From Munich’s Oktoberfest to Scotland’s whisky trail, why travelers are increasingly seeking food and drink at the source.

AI agents could trigger a bank run by automatically moving household cash from low-interest checking accounts to higher-yield alternatives, Slok warns.

Most people have hidden a spare key under a doormat at some point. It feels safe because nobody walks down the street checking every doormat.

AI agents do, and they never get bored. That shift in who does the checking helps explain why OpenAI has stopped training its most capable models for the second time in three months.

The company said it will resume training “only when we are confident that we have additional safeguards” in place, the Associated Press reported.

The pause came hours after OpenAI disclosed that its agents acted unexpectedly on several U.S. government websites this summer. The first halt followed July’s breach of AI platform Hugging Face by OpenAI agents.

The keys were already lying in plain sight

The government incidents share a detail that is easy to miss. At the Education Department, OpenAI agents found API developer keys to government data, according to the AP. At the Census Bureau, agents used login credentials they found online to pull public data, CNN reported.

That pattern runs through OpenAI’s own disclosures. The company’s incident tracker lists “use of exposed credentials” as a category of agent behavior, and it has notified dozens of affected third parties. A separate OpenAI report describes an internal model searching public GitHub repositories for leaked API keys during training.

The Hugging Face breach began the same way. On July 10, an agent found 14 publicly exposed credentials with write access and shared them with other agents, according to OpenAI’s technical account.

The agents misbehaved, but people left the doors unlocked. That makes this a security hygiene story, too, and no training pause fixes it.

OpenAI paused training of its most capable models for the second time in three months after its agents probed U.S. government websites.SOPA Images / Getty Images

Government sites attract agents because they are trusted

Washington did not land in this story by accident. An OpenAI spokesperson told CNN that “our models often turn to them as authoritative sources of public information.” The credibility that makes .gov pages useful also puts them on the front line of agent traffic.

Persistence turns that traffic into risk. OpenAI’s Hugging Face review found its agents seldom abandoned tasks, even impossible ones, and took riskier paths as they spent more effort.

Related: Google, OpenAI, and Anthropic just made a move on AI safety

Australian Prime Minister Anthony Albanese said an OpenAI agent that breached a Medicare statistics portal in June “didn’t accept no for an answer,” Reuters reported.

The latest trigger fits the same mold. On Sept. 20, an agent stuck on a search task used a DNS gap to query an outside chatbot, according to OpenAI’s incident report. An automatic shutdown failed, and staff stopped the run manually about 2.5 hours later, Fortune reported.

Federal damage so far appears limited. “No nonpublic information was accessed,” the SEC said, and the Education Department found no impact on its website or databases.

Pausing is becoming part of how OpenAI operates

OpenAI is private, so most investors hold it indirectly. Microsoft (MSFT) owned roughly 27% of OpenAI after its October 2025 restructuring, Bloomberg reported. That makes OpenAI’s development pace a direct input into how Wall Street values Microsoft’s OpenAI stake.

The delays carry a real price. OpenAI said its summer pause, including a hold on its largest planned training run, came “at significant cost and delays to frontier research.” Now the company says it expects to hit pause again as its models improve.

CEO Sam Altman told Fortune this month that “right now would be an ill-advised moment to go public,” ruling out a 2026 IPO. For future shareholders, safety stops now resemble weather delays at an airline: recurring, costly, and impossible to schedule.

More OpenAI:

Google, OpenAI, and Anthropic just made a move on AI safety

OpenAI makes development moves to counter SpaceX and Meta

Anthropic makes things more uncomfortable for OpenAI ahead of IPO

Voluntary brakes leave a gap Washington hasn’t filled

Federal policy is heading the other way. President Donald Trump agreed with Chinese President Xi Jinping this week to share information on AI dangers, the AP reported. Yet he told reporters the U.S. is not “putting on brakes.”

That leaves the brakes in private hands. OpenAI decides when to stop, what to disclose, and when to warn the people affected. Australia was not notified of its June breach until Sept. 10, and Albanese criticized the delay, according to CNBC.

Outside researchers are filling part of that gap. AI evaluator Transluce flagged an attempted hack on an Education Department site, which OpenAI has not confirmed. Transluce also found rogue activity on government sites in five states, some not clearly tied to OpenAI, CBS News reported.

That last detail widens the lens. With rivals Anthropic, Meta, and Google also reporting agents going rogue, does that mean going rogue is a new benchmark for measuring the capability of AI agents?

Much of today’s web security quietly assumes attackers are scarce and eventually give up. Agents break both assumptions at once. OpenAI can harden its own sandbox, but it cannot collect every key left on a public server.

Australia is now checking for more breaches, Reuters reported. The open question is whether Washington hunts for its own doormat keys before the next pause or after it.

Related: OpenAI’s agents breached Hugging Face. Nvidia wants it.

The competition to develop artificial intelligence models is intensifying.

Palantir Technologies (PLTR) seems to be getting more and more comfortable with that.

Instead of betting the firm on a single large language model, Palantir is building its artificial intelligence platform, or AIP, to work with models from competing AI companies while also connecting them to corporate data, permissions and real-world workflows.

That strategy became more obvious on Sept. 24.

Palantir has made xAI’s Grok 4.7 available to eligible AIP commercial customers. It also added on the same day, OpenAI’s GPT-6 Sol, GPT-6 Luna and GPT-6 Astra, and DeepSeek V4.1 Flash in some environments. In September, Palantir also added open-weight models from Z.ai and Moonshot AI, along with Google’s Gemini 3.8 Flash.

The collection is significant because Palantir doesn’t have to guess which AI lab will ultimately produce the most powerful model.

Its opportunity might be a layer higher: becoming the software companies use to work out which models can access their data, what they are allowed to do with it, and how AI gets dropped into real business operations.

This capability could become increasingly important as enterprises move from testing chatbots to deploying autonomous artificial intelligence systems.

Palantir wants AIP to sit above the AI model wars

Palantir has demonstrated how methodically it is growing the number of models that can run inside AIP with the September releases.

Grok 4.7 was added for eligible commercial environments with xAI enabled. OpenAI’s newest models are now available through OpenAI and Azure OpenAI integrations. Palantir also opened access to models from Google, Anthropic, DeepSeek, Z.ai, and Moonshot AI in several security environments.

This is a big deal.

A bank, manufacturer, or government agency may not want to rebuild its AI architecture each time a different model gets better at coding, reasoning, document analysis, or autonomous tasks.

Palantir is trying to make the model more replaceable while maintaining the customer’s underlying data, permissions, and operational structure.

Another September release builds on that idea.

Palantir announced general availability of AIP Evolve on Sept. 8. It orchestrates AI agents that can try to improve existing artificial intelligence systems. Goals can be reducing cost, reducing latency, improving evaluation scores, migrating workloads to different models etc. Then proposals can be reviewed before changes are merged into production, says Palantir.

Related: UBS resets Palantir stock price target for the rest of 2026

That means Palantir isn’t merely giving customers access to multiple AI models. It is also building software that helps decide when to change those models. CEO Alex Karp has made data control a central part of that pitch.

“Demand for AI sovereignty has now been unleashed,” Karp said when Palantir reported second-quarter results.

For Palantir, “sovereign AI” is only about clients preserving control of their unique data, models, infrastructure, and operational choices, instead of giving those benefits to an outside model supplier.

Now the approach is stretching well beyond Palantir’s own product.

Nvidia and Nebius deepen Palantir’s AI strategy

Palantir’s partnerships with Nvidia and Nebius are a glimpse into how the company wants this model-agnostic approach to work in practice.

Nvidia and Palantir announced a sovereign AI system for complex supply chains on Sept. 10, with an initial deployment inside Nvidia. By combining Nvidia’s open Nemotron models with Palantir Foundry, AIP and Palantir’s Ontology, supply chain information will be analyzed to help guide operational decisions.

More Palantir:

Palantir CEO admits AI would make him 20 times richer

Microsoft CEO adds fuel to Palantir CEO’s AI warning

Palantir CEO has a blunt verdict on OpenAI and Anthropic

The Nvidia deployment is notable because one of the companies at the heart of the AI infrastructure boom is using Palantir’s software.

Nvidia said it has built a digital supply-chain command center using Palantir Foundry. The chip giant is also training Nemotron models on operational decisions and the reasoning behind them in an effort to codify expertise that previously lived with human planners.

Palantir named Nebius (NBIS) as its preferred sovereign-AI infrastructure partner just two days ago.

Palantir says that once integrated, eligible commercial customers will be able to access Nebius compute and inference infrastructure from within Palantir’s enterprise perimeter, which could give customers more control over their compute, data, and AI models.

The companies are also looking to speed up the deployment of AI compute capacity, including modular data centers in areas that already have power.

Palantir’s smartest AI bet may be avoiding one big betJohn Lamparski / Getty Images

Palantir’s business growth gives the AI strategy more weight

Palantir’s AI strategy would be much easier to dismiss if it wasn’t showing up in the company’s financial results.

Second-quarter revenue surged 93% from a year earlier to $1.94 billion.

U.S. commercial revenue climbed 149% to $764 million, while U.S. government revenue increased 90% to $809 million. Total U.S. revenue reached $1.57 billion, up 115%.

Palantir also closed 220 deals worth at least $1 million during the quarter.

Of those, 98 were worth at least $5 million and 73 were worth at least $10 million.

U.S. commercial remaining deal value increased 124% to $6.24 billion, while U.S. commercial total contract value reached a record $2.13 billion, up 153%.

Profitability expanded alongside the growth.

Palantir generated $912 million of GAAP operating income, representing a 47% margin. Adjusted operating income was $1.19 billion, for a 62% margin.

The company subsequently raised its 2026 revenue forecast to between $8.15 billion and $8.158 billion and said U.S. commercial revenue should exceed $3.424 billion, representing growth of at least 134%.

Another major engine of growth is government spending.

On Sept. 17, the U.S. Army awarded Palantir USG a $48.1 million delivery order for software to replace nine legacy ammunition-management systems.

The initial term is 12 months and up to five optional additional years. The Army said the platform is intended to offer a single view of ammunition across planning, production, procurement, storage, distribution, and ultimate disposition.

That came just weeks after an Army milestone.

The Army has advanced the Tactical Intelligence Targeting Access Node, or TITAN, into production, with Palantir awarded $127 million in an initial eight-system production order.

Those awards together point to an unusual aspect of Palantir’s business.

Its rapidly growing commercial AI operation isn’t replacing its government business. Both are expanding at the same time.

Palantir’s biggest opportunity could depend on staying model-neutral

Valuation remains the big question for investors.

Palantir closed at $189.67 a share on Sept. 25. Wall Street remains broadly split on the value of the company’s unusually fast growth. Rosenblatt has a recent Buy and $225 target, while UBS reiterated a Buy and upped its target to $250 earlier this month. Other analysts are much more wary.

That disagreement shouldn’t be surprising.

Investors are trying to determine how much future AI adoption is priced into Palantir’s valuation, as the company posts revenue-growth rates rarely seen at its scale.

But do Palantir’s September product releases give investors another way to think about the company?

The best AI model might change again and again.

OpenAI might lead one category. Another category leader might be xAI, Google or Anthropic. Open-weight models could get a lot cheaper or a lot more powerful.

Palantir’s strategy seems increasingly constructed so that it does not necessarily have to know the winner beforehand.

If enterprises continue to adopt multiple models and want to maintain control of sensitive information and operational workflows, Palantir can try to sell the software layer that connects those pieces.

That may be the larger ambition behind what looked like a fairly simple product update on Sept. 24.

Palantir is not merely adding more AI models.

It is trying to make the identity of the winning AI model matter less to Palantir.

Related: Palantir faces a $330 million test U.S. investors should watch closely

The new software lets companies add custom security checks so they do not fall victim to the same single-point-of-failure vulnerabilities that plagued rival bridges.

CoinDesk found 27 successful swaps moving about 2,390 ETH into 75.2 BTC, even as Bitget urged THORChain to stop serving addresses tied to the $387.5 million theft.

While some chains have opted for locked shelves, Dollar General has taken a bolder approach to cutting down on shoplifting.

Essentially, the chain has made the decision that you can’t steal something if it’s not in the store.

Of course, that comes with a second clear caveat. You also can’t buy merchandise that’s not in the store.

“As we go forward, the team is looking at continued SKU rationalization, albeit probably more surgical in nature as we move forward. We’re already implementing some tests and learns even in the back half of this year around lower-volume store-type planograms, taking a substantial amount of SKUs out of the mix where it may not be as productive,” CEO Todd Vasos said during the chain’s second-quarter earnings call.

Removing lower-volume items to focus on faster-selling products makes sense, but that’s not the only reason Dollar General is removing items from its shelves.

“And also, again, in some high-shrink locations where shrink is still a bit of a headwind in some of these stores, and looking at SKU rationalization a little bit differently there,” he added.

Dollar General has fought a shrink problem

Shrink has been a Dollar General problem for years.

Vasos addressed it during the company’s fourth-quarter 2023 earnings call and shared some clear steps the company was taking to address the problem.

“Our second course of action will apply to all remaining stores with self-checkout, where we have begun limiting self-checkout to transactions consisting of five items or less. And finally, over the first half of the year, we plan to completely remove self-checkout from more than 300 of our highest shrink stores,” he said.

He framed the change as a positive.

“Collectively, we believe these steps are in line with where the customer wants us to be, which includes increasing personal engagement with them at the store. Additionally, we believe these actions have the potential to have a material and positive impact on shrink as we move into the back half of the year and into 2025,” he added.

Dollar General has removed self-checkout from thousands of stores.Shutterstock

Dollar General made aggressive shrink changes

One quarter later, Vasos addressed shrink again.

“Shrink continues to be the most significant headwind in our business, and we are deploying an end-to-end approach to shrink reduction across the organization, including efforts in our supply chain, merchandising, and within our stores,” he said during Dollar General’s Q1 2024 earnings call.

Dollar General’s shrink definition includes theft, but also damages and operational/administrative losses. The company’s own current results show that shrink and damages are being discussed separately.

“To help combat issues around shrink, our supply chain teams are primarily focused on ensuring deliveries are on time and in full and our merchants on reducing the amount of inventory we carry. Within our stores, we are focusing on delivering a more consistent front-end presence, broaden the reach of our high shrink planograms, which include the removal of high shrink SKUs and the elimination of self-checkout in the vast majority of stores,” he said.

That included a massive reduction in self-checkout.

“As we discussed on last quarter’s call, we converted approximately 9,000 stores away from self-checkout during the quarter. Following the quick and successful conversion of these stores in Q1 and given the ongoing challenge from shrink, we converted approximately 3,000 additional stores away from self-checkout in May, bringing us to approximately 12,000 conversions completed in total,” he added.

Dollar General has roughly 21,000 U.S. stores.

Shrink remains a major retail problem

While Dollar General is still making changes to fight shrink, Vasos did acknowledge that some of the chain’s efforts have worked.

“We were also pleased with the continued improvement in damages and shrink in Q2, which reflects strong in-store execution by the team,” he said during the Q2 2027 call.

Shrink, it should be noted, includes more than just shoplifting.

“Shrink (or shrinkage) is a measurement of inventory loss as a percentage of sales during a specific inventory period. It is used to forecast or account for losses in a retail balance sheet,” according to rhe National Retail Federation (NRF).

It’s a broad category that covers a number of different ways inventory disappears.

“Shrink calculations include losses stemming from theft (by employees and non-employees), administrative or operational errors, mistakes and other identified inventory loss. It is the most common form of measurement and benchmarking regarding retail loss. It also has its flaws,” the NRF added.

Still, shoplifting is an important part of the equation, and Capital One shared some 2025 data on that problem.

About half of retailers reported more shoplifting incidents in 2025 than in 2024.

In 2026, retailers are expected to lose $49.8 billion to retail theft. 

Projections indicate shoplifting could cost retailers more than $59 billion in 2029.

48% of retailers reported more shoplifting events involving individuals in 2025 than in 2024.

53% of retailers reported more incidents of team shoplifting (that is, multiple individuals stealing multiple items).

40% of retailers also reported fewer smash-and-grab events in 2025 compared to 2024.

Juveniles aged 12-16 are most likely to shoplift compared to other age groups.

Dollar General remains confident that it can continue to, well, shrink its shrink problem.

“We expect continued improvement in shrink and damages,” CFO Donny Lau said during the Q2 2027 call.

Vasos is confident in the future of the company.

“We feel we’re doing the exact right thing for the customer at the right time from a position of strength and on the offense. And we have the ability to flex up and flex down and have enough dry powder in the back half to be there for the customer every day,” he added.

Related: Coca-Cola, Pepsi killed 3 holiday soda flavors fans still miss

AI-driven traffic, plus deal-seeking consumers are expected to add up to a strong holiday season for online sales this year.

Alan Jackson’s “Where Were You (When the World Stopped Turning)” returns to the Digital Song Sales chart at a new career-high No. 11.

Your day-ahead look for Sept. 28, 2026