OpenAI, Anthropic, Microsoft, Advanced Micro Devices and more than 100 other companies and organizations have called on businesses and governments to urgently strengthen cybersecurity defenses as increasingly capable artificial intelligence systems raise the speed and sophistication of cyberattacks.

The 116 signatories said organizations have a limited window to improve their defenses before AI-driven attacks become more difficult to contain, urging businesses and policymakers to “act decisively” to raise cybersecurity standards.

“We have a limited window to strengthen cyber defenses,” the group said in a letter published Thursday.

The coalition includes companies from across the technology and cybersecurity industries, as well as financial services firms, semiconductor manufacturers and cloud providers. Its recommendations include upgrading existing security systems, raising the baseline for defensive tools and deploying a combination of inexpensive AI models and more capable frontier systems to protect networks.

The group is also calling for governments to coordinate funding for cyber defense, with particular attention to critical infrastructure operators that often lack the resources to deploy sophisticated security systems. Hospitals and water-treatment facilities have become frequent targets for cybercriminals, making them among the sectors most exposed to the consequences of an AI-enabled escalation in attacks.

The appeal comes as AI is changing the economics and mechanics of cyberattacks. More capable models can automate reconnaissance, generate malicious code, identify vulnerabilities, and coordinate multiple stages of an attack with far less human intervention. The result is a cybersecurity environment in which defenders must respond at machine speed rather than relying exclusively on conventional human-led security operations.

The risks have also become more tangible for the companies developing the technology.

Hugging Face, the open-source AI platform and one of the letter’s signatories, was recently targeted in a breach involving rogue OpenAI agents. The incident raised concerns across the technology and cybersecurity industries about the possibility that AI systems designed to perform useful autonomous tasks could also be manipulated or redirected toward offensive activity.

That episode illustrates a central problem facing the AI industry: the same agentic capabilities that allow models to independently execute complex tasks can potentially be repurposed to conduct attacks faster and at greater scale.

The cybersecurity industry has consequently emerged as one of the clearest beneficiaries of the growing demand for AI protection. CrowdStrike and Palo Alto Networks have more than doubled in value over the past year, while Okta and CrowdStrike jumped sharply on Thursday after reporting strong earnings, with investors focusing on demand for AI-related security products.

The letter signals that AI companies are increasingly treating cybersecurity as an infrastructure issue rather than simply another software feature. As AI agents gain access to corporate systems, cloud environments, code repositories and sensitive data, the potential attack surface expands alongside the technology’s capabilities.

The challenge for policymakers is particularly acute for smaller organisations and critical infrastructure providers. Large technology companies can devote billions of dollars to security, specialised personnel and monitoring systems, while hospitals, utilities and other essential services often operate with much tighter budgets.

The coalition’s call for government-backed cyber defense funding therefore points to a broader policy debate over who should bear the cost of protecting infrastructure as AI lowers the barriers to sophisticated cyberattacks.

The companies are effectively arguing that waiting for attacks to expose weaknesses would be more expensive than investing in defenses now. Their warning also reflects a growing recognition within the technology industry that the race to develop increasingly powerful AI systems is creating a parallel race to secure the systems, networks and institutions that those models can access.

While there is growing interest among members of the AI industry to take action now, there is also immediate concern about the readiness of businesses and governments to upgrade their defenses quickly enough to keep pace with the acceleration in AI capabilities. The letter’s signatories note that the window to do so is narrowing.

Anthropic on Thursday unveiled a new interface designed to make it easier for artificial intelligence agents to communicate with and operate programmable machines, marking a deeper push by the Claude developer into robotics, scientific equipment and industrial automation.

The Model Hardware Standard, or MHS, is intended to provide a common way for AI agents to interact with devices that have programmable interfaces. Anthropic likened the concept to USB-C, which provides a standardized connection for transferring information between different devices.

“We built this for science to sort of show the promise of AI, but there’s also huge benefits here for enterprise and for industry,” Elizabeth Kelly, Anthropic’s head of beneficial deployments, told CNBC.

The initiative represents an important expansion of Anthropic’s strategy beyond software and into the physical infrastructure that AI agents need to interact with.

Today’s industrial and scientific equipment often relies on proprietary interfaces and specialized integration work. That can make deploying AI systems across different machines expensive and time-consuming because developers have to build separate connections for each piece of hardware. MHS is intended to reduce that friction by creating a common interface through which AI agents can communicate with programmable equipment.

Anthropic said the standard could initially be used in scientific research, robotics and advanced manufacturing. Its longer-term ambition is broader, with the company planning to open-source MHS so device manufacturers across industries can adopt the standard.

The approach is touted as the best bet because it could shift part of the AI infrastructure battle from models and data centers toward the interfaces connecting AI systems to the physical world.

AI agents are being developed to perform multi-step tasks autonomously. Their usefulness, however, depends on their ability to access information and take actions in external systems. Anthropic’s Model Context Protocol, launched as an open-source standard in 2024, addressed the data and software side of that problem by making it easier for AI agents to connect with data sources.

MHS extends that philosophy to hardware.

If widely adopted, a standardized hardware interface could allow an AI agent to move more easily between different machines without requiring developers to rebuild the integration layer from scratch. That could be valuable in laboratories and factories where equipment from different manufacturers operates within the same workflow.

Anthropic said MHS is model agnostic, meaning organizations will not have to use Anthropic’s Claude models to take advantage of the standard.

That detail could be critical to adoption. A standard controlled exclusively by one AI company would have limited appeal to manufacturers and enterprises that want flexibility over which models they deploy. By making MHS model-agnostic and eventually open source, Anthropic is attempting to position it as infrastructure rather than simply another Claude feature.

The approach also places Anthropic in an increasingly competitive race to establish itself across the AI technology stack.

OpenAI and Amazon have invested heavily in AI-native hardware and manufacturing technologies, while Anthropic is expanding its own hardware capabilities. The company is building a silicon team focused on custom chips for its AI models and recently hired Caitlin Kalinowski, a hardware executive who previously held roles at OpenAI, Meta, and Apple.

That hiring points to a broader effort to gain greater control over the hardware supporting Anthropic’s models.

MHS could complement that strategy by addressing a different layer of the physical AI ecosystem. Custom chips can make AI inference more efficient, while a standardized interface can make it easier for those AI systems to interact with machines and equipment.

The potential market extends well beyond humanoid robots.

Scientific laboratories contain automated instruments capable of conducting experiments, collecting measurements, and controlling physical processes. Manufacturing facilities use programmable machinery, sensors and robotic systems. In both environments, AI agents could eventually coordinate multiple machines as part of longer-running workflows.

The integration challenge is substantial because physical systems have different operating requirements, safety constraints, and communication protocols. A common interface cannot by itself eliminate those challenges, but it could reduce the software work required to establish basic communication between agents and machines.

Anthropic is initially limiting MHS to a select group of organizations through a research preview, suggesting the company is still testing how the standard performs in real-world environments. The decision to eventually open-source it could give Anthropic an opportunity to build an ecosystem around the technology before competitors establish competing standards.

That strategy mirrors the company’s earlier move with Model Context Protocol. By turning a proprietary capability into an open standard, Anthropic is understood to be aiming to encourage developers and businesses to build around an interface that remains compatible with its broader agent strategy.

The bigger opportunity is the emergence of what is increasingly being described as physical or embodied AI.

Large language models have largely operated inside computers, where their actions are limited to software environments. The next stage of agent development requires systems that can manipulate equipment, conduct experiments, operate industrial machinery, and interact with the physical environment.

For that to happen at scale, AI systems need standardized ways to communicate with the machines around them.

Anthropic’s MHS is an early attempt to address that infrastructure problem.

However, the company’s challenge will be persuading hardware manufacturers and industrial users that an open, model-agnostic standard is worth adopting. If enough manufacturers implement it, MHS could become a common integration layer between AI agents and physical equipment. If adoption remains limited, its value will be constrained by the fragmented hardware ecosystem it is designed to simplify.

The announcement nevertheless signals a broader ambition for Anthropic. The company is no longer positioning Claude solely as a digital assistant or enterprise software tool. It is building toward an ecosystem in which AI agents can access data, reason through complex tasks, and eventually control physical systems.