The Information AM

Save 25% on an annual subscription to read the most important news about technology and business first. For even more, Save $250 on The Information Pro for unlimited access to our proprietary org charts, databases and surveys.

1.	Anthropic Says Its Models Also Hacked Outside Sites During Testing By Rocket Drew Source: The Information

Anthropic said its AI models hacked into three outside organizations during testing of their cybersecurity capabilities, incidents it discovered during a review prompted by a similar episode at rival OpenAI that has fueled widespread concern about AI safety.

In a blogpost Thursday, Anthropic said the direct impact of the incidents was limited, with the most significant episode involving extraction of some user credentials and several hundred rows of production data from an internal database. Anthropic said it successfully contacted two of the affected organizations on Monday and was still seeking to reach the third. It didn’t name the organizations.

The hacks occurred when Anthropic tasked its Claude models with solving a “capture the flag challenge,” a standard cybersecurity exercise that involves retrieving a piece of text hidden in computer systems.

In six of the 141,006 testing instances Anthropic reviewed, its models reached the internet because of a configuration mistake in the test environment. Four of those instances targeted the same organization. Claude didn’t use the internet access to make a copy of itself in any of the instances.

The hacks were conducted by three different Anthropic models, including Mythos, its highly cyber-capable model. In the Mythos attack, it uploaded malware to a public software registry, and 15 computer systems downloaded and ran the malicious software.

That behavior lines up with an earlier Anthropic finding while training Mythos that “the model occasionally circumvented network restrictions in its training environment to access the internet and download data that let it shortcut the assigned task.”

Anthropic said these hacks differed from OpenAI’s—in which that company’s models attacked the model repository Hugging Face—in that OpenAI’s models exploited a novel software vulnerability to access the internet, whereas Anthropic’s had internet access available (though they were instructed that they did not).

2.	OpenAI Slashes Prices on Some of Its Newest Models By Jason Dean Source: The Information

OpenAI said it is cutting prices on two of its newest models weeks after their release, as AI companies respond to customer concerns about surging bills for their services.

The changes, effective Thursday, affect two of the three GPT-5.6 models that OpenAI released broadly on July 9, the company said. OpenAI cut the application programming interface prices for Luna, the lowest-tier version, by 80% to $0.20 per million tokens for input and $1.20 per million output tokens. The API price for Terra, the mid-tier model, is falling 20% to $2 per million input tokens and $12 per million output tokens.

OpenAI didn’t change the API price of the most advanced GPT-5.6 model, Sol.

Business users have been contending with soaring costs for AI coding tools and other services, in part because Anthropic and other AI companies have shifted to usage-based billing. OpenAI had already been arguing that its new models are more efficient for customers to use, and Meta Platforms emphasized its lower prices when it launched its new Muse Spark 1.1 model with an API for business users early this month.

3.	AWS Revenue Growth Accelerates Sharply to 37% in Q2 By Catherine Perloff Source: The Information

Amazon stock surged in after hours trading after the ecommerce-and-cloud firm reported that revenue from its cloud division Amazon Web Services soared 37% in the second quarter, nine percentage points faster than the first quarter, bringing the division’s revenue to $42.2 billion. AWS’ operating margin also improved to 39% compared with 37% in the first quarter.

On the earnings call, Amazon said it would spend $220 billion in capital expenditures this year, a larger figure than the $200 billion estimate Amazon gave earlier this year. CEO Andy Jassy said that the higher cost of memory chips was driving this figure.

Buoying AWS’ growth is its AI business, which Amazon said now had an annual revenue run rate of $25 billion (although Amazon didn’t say exactly what went into that figure). Amazon’s chip business, which is the revenue it makes renting out Trainium and Graviton chips, also rose to an annual revenue run rate of $25 billion from $20 billion in the first quarter.

Earlier this month, Amazon laid off employees working on the company’s efforts to build its own frontier models. Jassy said on the call that Amazon can have a wildly successful business “without its own frontier models,” but that Amazon was still pursuing frontier model development to manage AI costs in Amazon’s own consumer applications and for its end customers. Amazon recently renegotiated its deal with Anthropic, such that using its models will be more expensive for Amazon.

Overall, Amazon’s revenue grew 20%. Its North American stores business grew 16%, thanks in part to Prime Day which was moved to the second quarter from the third quarter. The company said growth would slow to 9% to 12% in the third quarter.

4.	Exclusive: AWS Taps Apple Executive to Lead Key AI Products By Catherine Perloff and Kevin McLaughlin Source: The Information

Ori Herrnstadt, an engineering veteran of both Apple and Google, joined Amazon Web Services in May as vice president of compute AI services, an AWS spokesperson confirmed.

Herrnstadt’s role includes oversight of AgentCore, a service launched last October that customers use to build, run, and manage AI agents; and SageMaker, an AWS service for training and running AI models, the spokesperson said.

The hire marks a shift in product management duties at AWS. The vice president who previously led AgentCore, David Richardson, left in April after just eight months in the role. His manager, Swami Sivasubramanian, vice president of agentic AI, took on an expanded role this week as a strategic advisor to other teams developing AI products.

Sivasubramanian will continue leading products like Amazon Quick, an AI work assistant, and Kiro, an AI coding tool.

“We welcome Ori to AWS. The caliber of leaders choosing to build their next chapter at AWS reflects our significant momentum in AI and the enormous opportunity in front of us,” the AWS spokesperson said in a statement via email.

Herrnstadt joined Apple in 2015 as a cloud engineer and worked on the “privacy-enforcing data and compute platform that manages all user data,” according to his LinkedIn profile. He left in 2020 to join Google, where he worked in its core machine learning engineering team, and returned to Apple two years later as a senior director. The nature of his recent role at Apple couldn’t be learned.