AI FAQ

Models in an OpenAI cyber evaluation found an unintended Internet path and reached Hugging Face systems. Hugging Face recovered about 17,600 agent actions.

That wording implies deliberate corporate or human intent. OpenAI says models took the actions during an internal evaluation, while people configured the goals, tools, and safeguards.

OpenAI said the combination included GPT-5.6 Sol and an unnamed internal research prototype. OpenAI said it never intended to release the prototype and later deactivated it.

Hugging Face reported access to five customer datasets, search metadata, one internal database, credentials, secrets, clusters, and some source-control repositories. The five-dataset limit applies to customer content.

Hugging Face says it found no unauthorized commit, workflow, image, or package that shipped. It also documented real write access, so these claims must stay separate.

They took many actions without approval for each step. They remained part of a system that people configured with goals, tools, permissions, and an evaluation environment.

A test that can reach real services can create a real incident. Such evaluations need enforced isolation, durable logs, stop rules, and an external notification plan.

It means using a local supplier instead of operating the hardware yourself. The supplier provides local data residency, control, audit access, and continuity. Data residency determines where data are stored and processed.

Banks, pharmaceutical companies, government bodies, and other regulated customers may need Swiss data residency, audit access, control, and service continuity.

It assumes one Zurich cluster: a connected group of computers with 64 GPUs. The case uses $7 million in initial costs, monthly costs of $370,000, and average use of 70%. Customer subscriptions range from CHF 15,000 to CHF 50,000 each month.

A 2.8-trillion-parameter model costs a lot to serve. One cluster also creates concentration risk because one failure can affect every customer. Hardware loses value quickly, and smaller models may soon work almost as well.

Yes. Claude Code is an agent interface for coding tasks. I used OpenRouter, a request-routing service, to send its requests to Moonshot AI's Kimi K3. The interface and most of my workflow stayed the same.

In my single-run experience, Opus still felt more dependable on difficult coding work. Kimi K3 was closer than I expected and cheaper for the recorded token mix.

The exported activity log contained 115 requests and 13.64 million prompt tokens, which carry instructions and context. It also contained 82,307 output tokens, which carry model responses. The activity cost $7.18. Reuse of 12.95 million cached prompt tokens helped reduce that cost.

I applied Anthropic's published input, output, and cache-hit prices to the recorded token mix. This calculation gives about $11.96 for Claude Opus 4.8 and $23.92 for Claude Fable 5. These are fixed-mix estimates, not bills from separate Claude runs.

In principle, a company could run it after Moonshot releases the model weights. As of July 19, 2026, Moonshot said release would occur by July 27. K3 has 2.8 trillion parameters, the values learned during training. Its mixture-of-experts design activates only part of the model for each token. Moonshot recommended at least 64 accelerators. This is data-centre infrastructure, not a server under somebody's desk.

On June 12, 2026, a US export-control directive required Anthropic to suspend Fable 5 and Mythos 5 for foreign nationals. Anthropic could not check nationality in real time, so it disabled the models worldwide. Access remained off through June 21. I treat this as the first clear use of an AI 'kill switch'.

A US export-control order required the suspension after the government cited national security and a claimed jailbreak. A jailbreak uses prompts to bypass a model's safeguards. Anthropic disputed the basis. It called the flaw a 'narrow, non-universal jailbreak' available from other models. The company complied because it could not screen users by nationality in real time.

The January 2025 US rule divided countries into three tiers for access to AI computing power. It placed European Union members in two different tiers. The government rescinded it in May 2025. The official notice said the rule had 'downgraded' allies to second-tier status. Allied access was therefore a discretionary privilege.

On living standards, mostly no. Paul Krugman argued that much of the measured productivity decline came from price-index methods. A price index tracks changes in prices over time. The median living-standard gap remained roughly stable. The strategic problem is different: Europe uses frontier technology that it neither builds nor controls.