Hundreds of AI researchers call for stronger sandboxes to contain cyber-capable models

By NewSunSEO | October 15, 2026

More than 400 AI researchers have signed an open letter urging labs that build cyber-capable AI models to lock those systems down with high-assurance sandboxes before more security failures reach critical infrastructure. The coalition, which includes several of the field’s most prominent figures, frames a string of recent breaches at AI labs as preventable lapses in basic cybersecurity, and lays out a three-part plan for fixing them.

The letter, published on a Tuesday, calls on AI and security researchers to collaborate on alignment and containment. It argues that the incidents disclosed in recent months at several major labs already serve as evidence of why containment work matters, and warns that future failures on exposed infrastructure will be far more damaging if containment is not tightened now.

What the letter recommends

The letter sets out three concrete steps for the industry:

• Organizations developing and deploying cyber-capable AI models should immediately put high-assurance containment systems in place, and should start by defaulting to airgapping those models, meaning isolating them from unsecured external networks.
• An international standards body should be created, bringing together governments and a wide range of AI and security experts, to set best practices for alignment and containment and to define how controlled access outside the sandbox can be granted safely. The same body would also track every reported containment failure.
• Every security incident involving an AI containment failure must be reported to that standards body and made public, so the industry can learn from mistakes and be held accountable for them.

Which AI labs have been involved in recent breaches

The letter does not name any single company. It does, however, arrive in the wake of a July breach of Hugging Face linked to OpenAI and a series of incidents at several major AI labs that have since come to light. The letter’s authors argue these events show why containment work is urgent, and write that without immediate changes in how AI is developed and deployed, similar incidents will keep happening and future ones will hit harder.

Where the broader push for guardrails stands

The letter adds to a growing chorus from inside major AI companies. In recent months, leaders at Microsoft, former leadership at DeepMind, and Anthropic have publicly urged stronger security standards for advanced AI. At the same time, Reuters reported on the same Tuesday that Anthropic and OpenAI have been seeking influence over what third-party evaluators can publish, a push the report frames as aimed at protecting intellectual property but which researchers warn could also narrow the transparency and usefulness of outside reviews.

Why the timing matters for containment work

Containment work is gaining urgency because labs are racing toward AI systems that can improve themselves. That capability, sometimes called recursive self-improvement, or RSI, would shrink the window in which containment and alignment work has to be solved. A co-founder of Anthropic wrote in a Monday newsletter that full RSI could arrive in 2027, two years earlier than a previous forecast, which makes getting control of these systems right before they reach that point more important than ever.

What counts as a strong sandbox for an AI model

In the letter’s framing, a sandbox is a containment boundary around a model that controls what it can reach on the network, what tools it can call, and what data it can read or write. Airgapping is the strictest version: the model runs on systems with no connection to unsecured external networks at all, so even if the model is misused or compromised, it has no live route out. High-assurance containment is the engineering discipline of building those boundaries so they hold up under deliberate attack, with verified components, audited code paths, and tested failure modes, rather than relying on ordinary production security that may not be designed for an adversary that is itself a capable AI.

FAQ

What are AI researchers asking for in the open letter on sandboxes?

They are asking labs developing cyber-capable AI models to deploy high-assurance containment systems, to airgap those models by default, and to back an international standards body that tracks containment failures and requires public disclosure of incidents.

Which AI security incidents prompted the letter?

The letter points to recent containment failures at several major labs, including a July breach of Hugging Face linked to OpenAI, and argues these incidents show why stronger sandboxing is needed.

Why are sandboxes more urgent now for AI labs?

Cyber-capable AI is being deployed even as labs race to develop systems that can iteratively improve themselves. If recursive self-improvement arrives, the window for establishing containment narrows as system capabilities increase.