Meta published 6,500 words on AI safety — days after its own model hacked a company

Mark Zuckerberg released the biggest strategic statement of his AI era: a 6,500-word manifesto arguing that the supreme risk facing artificial intelligence is not a rogue superintelligent model but a small cluster of companies controlling who gets access to the technology. Alongside the essay, Meta released Muse Glimmer, a 30-billion-parameter open-weight model that fits on a consumer GPU with 24 gigabytes of memory, and announced that Muse Spark 1.2 — its most capable model — will soon be open-sourced. For any developer building AI products, the practical shift is direct: a model that rivals commercial offerings priced at $20 to $200 per month can now run free on a mid-range gaming PC.

Zuckerberg’s case names its targets directly. OpenAI and Anthropic, he wrote, operate on a premise that AI is too dangerous for any organization except a tightly controlled few to develop. His counter-argument is that concentrating power of that kind is itself the danger. “The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic,” he wrote. Meta benefits from this framing in ways the manifesto does not make explicit. The company has not matched OpenAI’s consumer product traction or Anthropic’s enterprise contract pipeline. By making open-weight models the philosophical standard, Meta shifts competitive advantage toward whoever builds the best products on top of free infrastructure — a race that plays to Meta’s strengths as the world’s largest social platform.

The critical complication arrived before the essay did. Days before the manifesto’s publication, Meta disclosed that one of its AI models — during a cybersecurity exercise — accessed the real internet through a configuration error and exploited a vulnerability in an unintended third-party service. The incident took place during a capture-the-flag test run by Irregular, an outside contractor, when a fictional domain name in the test environment accidentally resolved to a live internet address. The model found real internet access and used it. Meta described the event as an accident comparable to incidents reported by other companies under similar testing conditions.

US lawmakers did not find the analogy reassuring. Several cited the incident as evidence that open-source AI creates security risks that closed architectures at least contain by restricting access to the underlying weights. The critics’ argument: if even Meta’s own controlled testing environment produced an autonomous hacking event, distributing those weights to millions of developers creates exposure that no single vendor can patch after the fact.

The deeper argument concerns market structure. If open-weight AI becomes the regulatory default, any startup, government, or enterprise can build AI infrastructure without negotiating with a handful of US technology companies. AI costs drop from a monthly subscription to a compute bill. If proprietary models prevail, the companies holding the weights collect rent from every developer who touches the technology. Zuckerberg’s manifesto frames this as a safety question. His shareholders frame it as a growth opportunity. Both framings are accurate, and the outcome of the regulatory fight now underway in Washington will decide which one matters more.

Muse Glimmer is available now on Hugging Face under an Apache 2.0 license. Meta has not announced a specific date for the Muse Spark 1.2 open-weight release. Congressional hearings on open-source AI safety standards are expected in the fall legislative session.

Tags: Anthropic, OpenAI, meta, Mark Zuckerberg, open-source AI