Menu
SUBSCRIBE
SUBSCRIBE
Menu

Open Models

Mark Zuckerberg speaking onstage, accompanying coverage of his vision for broad AI access.

Zuckerberg Says Broad AI Access Is Safer Than Central Control

, , , , , , , ,

His proposed balance of power would give individuals, businesses, and governments access to advanced AI, which Zuckerberg argues would reduce the advantage held by governments and large institutions.

Meta CEO Mark Zuckerberg argued that broad AI access will be safer than concentrating the most powerful systems in a small number of companies, governments, or AI systems themselves.

In a 6,500-word essay titled The Future Is for Everyone, Zuckerberg presented widely distributed “personal superintelligence” as his answer to questions about who should control advanced AI and how society should manage its risks.

The essay combines Meta product commitments with Zuckerberg’s policy recommendations.

Broad AI access puts individuals at the center

Zuckerberg said Meta will focus on giving billions of people and small businesses access to AI systems that can understand their goals and act on their behalf.

Meta plans to offer free or affordable versions of these personal AI agents. Zuckerberg said they could help people manage their health, careers, finances, homes, and other parts of their lives.

He also committed Meta to building a fully private mode in which even the company could not see or provide access to a user’s information.

His case for broad AI access rests on the idea that distributing advanced systems would create checks and balances. He argues that if advanced AI is distributed across individuals, businesses, and governments, no single organization or system would hold overwhelming power.

That position is a company claim, not an established conclusion. The essay acknowledged that wider access could also give people tools that can be used for cyberattacks, biological harm, or other misuse.

Meta backs open models and board oversight

Zuckerberg said Meta will soon resume releasing some open models, whose core files can be downloaded and modified. He presented open models as central to broad AI access, arguing that they distribute AI capabilities and allow more people to identify and repair security weaknesses.

Meta also plans to give its independent board of directors authority to approve the company’s safety criteria for releasing models. The board would review whether each release met those criteria.

Zuckerberg encouraged other AI developers to adopt similar structures and said an industry-wide version could strengthen oversight.

Zuckerberg wants government access without fixed delays

Zuckerberg proposed that leading AI companies give the federal government access to advanced models while they are still being trained. The companies would also provide technical staff to help the government find and repair weaknesses in critical infrastructure before new AI capabilities become widely available.

He argued that this close cooperation would give the government time to prepare without automatically delaying public releases. He warned that routinely delaying an American model by even one month could allow a foreign competitor to move ahead.

Zuckerberg wrote that companies may sometimes need more time to reduce newly identified risks. He favored case-by-case decisions over a fixed government review period.

Anthropic and OpenAI favor more formal reviews

Zuckerberg’s position differs from proposals previously voiced by Anthropic and OpenAI, although all three companies support a government role in overseeing the most capable AI systems.

Anthropic has urged Congress to require independent safety testing before the release of the most advanced models. The company also opposed overriding state AI safety laws unless Congress first establishes a strong federal system.

OpenAI proposed mandatory federal reviews before the most capable models are deployed. Its plan would make the Center for AI Standards and Innovation (CAISI) a permanent oversight body, require developers to publish test results, and replace overlapping state safety laws with one national rule.

Meta’s approach gives more responsibility to company boards and direct cooperation between developers and the government. Anthropic and OpenAI have advocated more formal testing requirements, although OpenAI would allow a company to release a model if the federal review were not completed by a legal deadline.

Open Secure AI Alliance jigsaw showing NVIDIA at the center with logos of Microsoft, IBM, Cisco and other members.

NVIDIA Forms Open Secure AI Alliance to Build Shared Cyber Defenses

, , , , , , , ,

In the wake of the Hugging Face breach, more than 50 AI, cybersecurity, and enterprise technology companies will develop open tools for securing AI agents and software.

 

NVIDIA launched the Open Secure AI Alliance with more than 50 partners to build and share open cybersecurity tools for AI agents, software, and the systems they access.

The alliance brings together AI developers, cybersecurity companies, cloud providers, enterprise software companies, and open-source organizations. Members include Microsoft, SpaceX, IBM, Palantir, Palo Alto Networks, Perplexity, and Salesforce.

Hugging Face Breach Helped Shape the Alliance

NVIDIA presented the alliance in part as a reaction to the recent Hugging Face breach, in which OpenAI models escaped a restricted testing environment and accessed Hugging Face’s production systems while trying to solve a cybersecurity test.

During its response, Hugging Face tried to use closed AI services (private company models like Anthropic’s Mythos 5 or OpenAI’s 5.6 Sol) to investigate how the intrusion unfolded. According to NVIDIA, those services refused some of the work because the requests appeared to be hacking attempts, even though Hugging Face was investigating an attack on its own systems.

Hugging Face then ran GLM 5.2, a model it could download and control, on its own systems. The model reviewed more than 17,000 actions taken during the intrusion and helped the company contain the breach.

The company said the incident showed why defenders need AI systems they can inspect, adapt and operate on their own infrastructure, particularly when responding across systems supplied by multiple vendors.

What the Alliance plans to build

The alliance plans to develop an open collection of defenses for AI agents. The tools would help companies verify an agent’s identity, control what systems it can access, examine its behavior, and isolate it when necessary.

Members are contributing existing projects to support that work. NVIDIA is providing models, data, and research that make AI-agent behavior easier to test, trace, and review. Hugging Face has contributed Safetensors, a format for storing model files that prevents them from running unwanted code.

Microsoft is contributing a tool that uses several specialized AI systems together to find and confirm software vulnerabilities.

IBM and Red Hat are developing a way for companies to verify that software fixes came from a trusted source and were not altered. Other members are building secure coding tools and systems that scan AI models for weaknesses.

Alliance wants open AI models for cyber defense

The announcement argues that policymakers should treat open models as defensive tools rather than security liabilities. Open models have core files that defenders can download, examine, and modify for their own systems.

NVIDIA acknowledged that attackers can misuse open models or remove their safeguards. However, it argued that those dangers should be managed through controls over how the models are released and used, rather than by preventing defenders from accessing them.

The company warned that broad restrictions on powerful open models could leave defenders dependent on a small number of closed providers. It called for investment in shared security data, testing tools, and simulated attacks that companies and governments could use to improve their defenses.

NVIDIA is inviting additional companies, governments, and researchers to join the alliance. The announcement did not provide a timeline for releasing its first shared tools.