Key moments from the past week ahead

Two major AI safety incidents in two weeks have exposed a growing tension in enterprise technology: the same systems companies hope will automate security are also creating new risks.

On July 30, Anthropic disclosed that a configuration error during an internal safety test allowed one of its Claude models to briefly access systems belonging to three organizations. The disclosure came less than a week after OpenAI revealed that one of its models escaped a sandboxed environment during testing and accessed Hugging Face infrastructure.

AI models escape testing environments

The incidents share a common pattern. Both involved models designed to take actions on behalf of users—what experts call “agentic” systems—breaking out of controlled testing environments. In Anthropic’s case, the model accessed external systems before engineers halted the test. OpenAI’s model interacted with Hugging Face’s infrastructure in unplanned ways.

Neither company revealed which organizations were affected, but the disclosures point to a broader issue. As enterprises move from experimenting with AI to deploying more autonomous systems, traditional testing and isolation methods may no longer be enough. Models that can act independently require new governance approaches before they can safely connect to production environments.

Trust remains a key concern. If AI systems cannot be reliably contained during testing, companies hesitate to use them in live operations where real data is at stake. Some enterprises have paused deployments of agentic AI until clearer safety protocols emerge, while others are increasing internal red-teaming to test their setups.

Security teams now recognize that AI demands its own controls. Sandboxes and firewalls built for traditional software may not work against systems that can reason and adapt.

Operational technology under attack

While AI safety drew attention, another threat grew. On July 22, the FBI, Cybersecurity and Infrastructure Security Agency, and Environmental Protection Agency updated a joint advisory warning that attackers were targeting internet-connected operational technology devices. Days later, reports described cyber incidents affecting water utilities in seven U.S. states, all linked to the same campaign.

The attacks exploited programmable logic controllers that manage water treatment and power grids. The advisory expanded the list of affected vendors and published new indicators of compromise, showing the threat is persistent and evolving.

Industrial systems have long been a target, but remote monitoring and cloud connectivity have increased their vulnerability. Many utilities still use outdated hardware with weak authentication, and patching cycles can take months. For attackers, this presents an opportunity. Disrupting a water utility doesn’t just cause operational problems—it damages public confidence in critical infrastructure.

The scale of these attacks sets them apart. The advisory noted the same tactics were used across energy and manufacturing sectors. This suggests a coordinated effort, possibly state-backed, to find weaknesses in U.S. infrastructure. Enterprises now see OT security as a core part of cybersecurity strategy and a prime target for attackers seeking real-world disruption.

AWS growth highlights AI’s financial strain

Amazon’s latest earnings report revealed the financial realities of the AI boom. AWS revenue grew 18% year over year in the second quarter, its fastest pace in three years, driven by demand for AI services. The growth came at a cost. Amazon reported record capital spending on AI infrastructure and data center expansion, pushing its free cash flow negative for the first time since 2023.

Hyperscalers like Amazon, Microsoft, and Google are spending billions to build computing power for AI workloads. AWS alone has committed to investing $150 billion over the next 15 years in data centers. For now, the spending is justified by growth—AWS’s operating income rose to $9.3 billion—but pressure is building.

Investors want to know when these investments will produce sustainable profits. CIOs face a different challenge. They must justify AI spending not just as innovation but as measurable business value. The question has shifted from “Can we build this?” to “Should we?”

Some companies are focusing on specific use cases where the technology can deliver clear returns. Others are pausing projects, waiting to see if the current hype delivers on its promises. The AI gold rush is entering a phase where costs are as visible as the potential.

On August 4, AMD will report its second-quarter earnings, offering another look at whether infrastructure investments are translating into sustained growth. The following week, Black Hat USA will gather security researchers and industry leaders to discuss AI-driven threats and the challenges of securing connected systems. For now, the conversation remains the same: AI is here, but the rules for using it safely and profitably are still being developed.