Prince Mario-Max Schaumburg-Lippe: Three Rival AI Labs Sit Down Together to Write Safety Rules

Competitors in the AI race do not usually share a table. This week, the three biggest names in frontier AI confirmed they have been doing exactly that. OpenAI, Anthropic and Google DeepMind have spent several weeks coordinating on AI safety work, and the collaboration became public as the labs jointly pointed to cyber-focused models, shared safeguards and access programs built to be used across company lines.

Bloomberg reported OpenAI’s confirmation that it is working with Anthropic and Google on safety, the latest step in a courtship that has been building since July, when Google DeepMind chief Demis Hassabis proposed an industry standards body modeled loosely on the way Wall Street polices its own brokers. Anthropic’s Dario Amodei added momentum in September with an essay arguing the frontier should be paced so safety work can keep up, and OpenAI’s Sam Altman publicly backed the idea.

What rivals can share without giving away the store

The smart part of this arrangement is its narrow scope. The labs are not pooling model weights or training recipes. They are coordinating on the defensive side of the most dual-use skill their models have: writing and analyzing code. A model that can find flaws in software can help a security team patch them, and the same skill, shared as a common evaluation problem, lets all three labs measure the capability the same way and build safeguards that speak the same language.

Shared evaluations matter more than they sound. Today, each lab grades its own homework with its own tests, which makes outside comparison nearly impossible. Common benchmarks, pre-release testing frameworks and a shared vocabulary for reporting incidents would give researchers, customers and regulators something the field has never really had: numbers that line up across companies.

Cooperation is becoming a pattern

Readers of this page have watched the industry’s center of gravity shift from solo launches to shared infrastructure all month. Yesterday we covered Anthropic putting 5 percent of its shares into a public benefit fund, a governance experiment, and last week Anthropic was already consulting law firms on how its agents should negotiate. The safety coordination fits the same arc: the labs are starting to build the institutions around the technology, not just the technology.

There is self-interest in it, and that is fine. A serious incident at any one lab would damage trust in all of them, so a shared floor for safeguards protects everyone’s business. The useful test is whether the coordination produces public artifacts – published benchmarks, joint evaluation results, common incident reports – rather than warm statements. The labs say access programs for vetted defenders are part of the package, which would be a concrete deliverable to watch for.

What to watch next

Three signals will tell whether this is substance or choreography. First, whether the proposed standards body gets a name, a budget and a launch date; reporting since September has pointed to late 2026 or early 2027. Second, whether Microsoft and other major labs join or stay with parallel efforts. Third, whether the first joint evaluation results are published with enough detail for outsiders to check.

For now, the headline is simple and, in this industry, genuinely unusual: the three firms racing hardest to build the most powerful AI systems on Earth have decided that some problems are better solved together. Given what their models can already do, that instinct arriving this early is the most encouraging safety news of the year.

The summit trail behind the handshake

The coordination did not appear from nowhere. This week also brought a royal-hosted AI safety summit in Scotland, where executives from the same labs sat down with Nvidia and others to talk about shared defenses. Put the two events together and a picture forms: an industry that spent years treating safety as a private, competitive discipline is quietly reclassifying parts of it as shared infrastructure, like aviation safety reporting or banking fraud databases. Nobody competes on whether the plane lands. The labs seem to be deciding that some parts of AI should work the same way.

There is a practical benefit for ordinary users in all this. Shared evaluations make it harder for any single company to grade a risky capability as safe by choosing a friendly test. If the three labs converge on common cyber benchmarks, a model that scores well in one house will have earned it in all three – and a model that scores badly will be much harder to ship quietly. That is a better deal for every customer who has to decide which system to trust with real work.

One more reason to take the coordination seriously: it survived contact with competition. These are the same three companies that spent the very same weeks fighting over enterprise contracts, research talent and, as our ai-5 story today details, the largest pools of capital in technology history. Cooperation that holds while the rivalry intensifies is not a photo opportunity. It is a judgment, made independently in three boardrooms, that the risks of uncoordinated capability growth are the one bill none of them wants to pay alone. History will grade the follow-through. The instinct, at least, is exactly right.

Also published on: