OpenAI, Anthropic, and Google logos side by side, symbolizing former AI rivals uniting against adversarial distillation
Voltar ao Blog
🇧🇷Português8 de abril de 20267743

Why Nobody Told You America's Biggest AI Rivals Just Became Allies

OpenAI, Anthropic, and Google put years of rivalry aside to fight Chinese adversarial distillation.

Publicado em 08/04/2026

Tags: adversarial distillation, OpenAI, Anthropic, Google, AI security, Frontier Model Forum, national security, AI theft, DeepSeek, Moonshot AI

Stop thinking OpenAI, Google, and Anthropic are just competing for your subscription money. For three years, these companies fought over everything: talent, compute, benchmarks, government contracts, and the public narrative of who was building the future. They cooperated on almost nothing. Then something happened that was serious enough to force all three of them into the same room. And what you haven't been told is exactly how bad the problem has gotten — or what it means for the future of AI that you use every day.

You were told the AI race was between the US and China. You were not told that Chinese competitors had been quietly stealing the results of billions of dollars of US AI research — systematically, at scale, for years — while the American labs watched it happen and couldn't agree on how to stop it. That changed on April 6, 2026. Bloomberg broke the story: OpenAI, Anthropic, and Google have begun sharing threat intelligence through the Frontier Model Forum — a nonprofit they co-founded with Microsoft in 2023 — specifically to detect and stop what experts call "adversarial distillation." The rivalry is over. The alliance has begun. Here's why this matters more than any model release you've seen this year.

What Is "Adversarial Distillation" — The Theft Nobody Calls Theft

To understand why three of the world's most competitive companies just became partners, you need to understand the specific attack they are defending against. Distillation is a legitimate AI technique. It works like this: you take a powerful "teacher" model, feed it thousands of questions, collect its responses, and use those responses to train a cheaper, faster "student" model that mimics the teacher's behavior. When companies create smaller, efficient versions of their own models — like a lightweight Claude or a compact Gemini — they use distillation. This is normal, accepted, and often beneficial.

Adversarial distillation is something different entirely. In adversarial distillation, an external competitor — one with no license, no agreement, and no legitimate access — bombards a frontier model with millions of automated queries, designed not to get answers but to extract the model's underlying intelligence. The goal is to train a copycat model that replicates years of safety research, reasoning capabilities, and fine-tuning at a fraction of the original cost — without investing a single dollar in original training. US authorities estimate this practice costs American AI labs billions of dollars in lost revenue each year.

The Attacks Already Documented — Bigger Than You Were Told

The coalition announced this week is not a response to a hypothetical threat. It is a response to documented, specific, ongoing attacks. In February 2026, Anthropic published a formal report identifying three Chinese AI laboratories by name: DeepSeek, Moonshot AI, and MiniMax. All three had been caught extracting capabilities from Claude — Anthropic's flagship model — and then stripping away the safety guardrails that Anthropic spent years developing. That last detail is not a footnote. It is the most dangerous part of the story.

The distilled models that emerge from these attacks don't just copy Claude's reasoning. They copy it without the safeguards. The result is a model that is nearly as capable as Claude at answering complex questions — but has no guardrails against being used for cyberattacks, disinformation, or weapons development.

One month earlier, in March 2026, Anthropic had identified a coordinated campaign involving 24,000 fake accounts specifically created to extract intelligence from Claude at scale. The accounts were systematically querying the model with high-volume, automated prompts designed to reconstruct its behavior from outputs alone.

OpenAI's experience was parallel. In a formal memo to the US House Select Committee on the Chinese Communist Party, OpenAI warned that DeepSeek had been using "increasingly sophisticated methods" to extract data from its models — methods that evolved and adapted every time OpenAI tightened its defenses. OpenAI accused DeepSeek of attempting to "free-ride on the capabilities developed by OpenAI and other US frontier labs."

Google's Threat Intelligence Group identified more than 100,000 prompts targeting Gemini's reasoning capabilities in a coordinated model extraction campaign. Three companies. Three separate investigations. The same attackers.

Why They Finally Decided to Work Together

For three years, the information-sharing that is now happening through the Frontier Model Forum simply did not happen. These companies considered each other competitors. Sharing vulnerability data with a rival felt like sharing trade secrets with an enemy. The decision to cooperate is modeled on a practice the cybersecurity industry built over decades: Information Sharing and Analysis Centers, where companies in the same sector exchange indicators of compromise — attack signatures, IP addresses, behavioral patterns — so that when one company is hit, the others can immediately strengthen their defenses.

When Anthropic detects a novel distillation pattern — a new technique for extracting capabilities from Claude — it now flags that pattern to OpenAI and Google. When Google identifies a new botnet of fake accounts hammering Gemini, that information flows to the other labs before the same botnet can target ChatGPT or Claude. The three companies have also begun developing shared technical countermeasures: technologies to identify abnormal traffic patterns, high-volume coordinated queries, and behavioral signatures of automated cloning attempts. When attacks are confirmed, they can respond collectively — banning accounts, blocking IP ranges, altering rate limits, and modifying output formats in ways that degrade the quality of data available for distillation.

The National Security Dimension Nobody Is Discussing Enough

Here is the aspect of this story that the technology press has underreported. Anthropic stated publicly that the distillation threat "extends beyond any single company or region" and poses a national security risk. This is not corporate hyperbole. It is a technical reality with specific consequences.

When a safety-stripped version of Claude or GPT-5.4 reaches the hands of state actors, criminal organizations, or terrorist groups who cannot access the original models due to export controls and terms of service — the capability gap that those safety measures were designed to create disappears. Every hour of safety research that Anthropic, OpenAI, and Google have invested — the red-teaming, the constitutional AI training, the RLHF fine-tuning designed to prevent the models from helping build bioweapons or conducting cyberattacks — can be stripped away in hours by adversarial distillation. The result is a model almost as capable as the original, with none of the protections. This is why OpenAI sent a formal memo to Congress. This is why Anthropic banned all Chinese-controlled companies from using Claude. This is why Google published a public blog post about the threat. These are not competitive posturing moves. They are security disclosures.

What This Alliance Means for the Future of AI

The formation of this intelligence-sharing coalition represents something historically significant: the world's most powerful AI labs acknowledging that the threat to safe AI development is greater than their competitive differences. When OpenAI, Anthropic, and Google decide that the risk of Chinese adversarial distillation is serious enough to share their most sensitive security intelligence with each other — companies that compete for the same enterprise contracts and the same research talent and the same government relationships — the scale of the problem they are responding to is enormous.

But here is the part that should inspire rather than frighten you. This coalition is also a signal that the institutions of safe AI development are maturing. The frontier labs are not just racing to build more powerful models. They are building the infrastructure — the threat intelligence networks, the shared defense protocols, the policy engagements with Congress — that will protect the integrity of AI for everyone who uses it. The future of AI is not a race between companies. It is a collective project. And three of the most important actors in that project just decided to act like it.

Want to follow every major AI move the moment it happens? Join thousands of readers at fabioraphaelli.com and get the biggest AI news every week.

Frequently Asked Questions

What is adversarial distillation in AI?

Adversarial distillation is a technique where a competitor bombards a frontier AI model with millions of automated queries, collects the outputs, and uses them to train a cheaper copycat model that replicates the original's capabilities — without licensing, authorization, or investment in original training. US authorities estimate this costs American AI labs billions of dollars in lost revenue annually.

Why did OpenAI, Google, and Anthropic start sharing intelligence?

On April 6, 2026, Bloomberg reported that the three companies began sharing threat intelligence through the Frontier Model Forum — a nonprofit they co-founded with Microsoft in 2023 — to detect and stop adversarial distillation attacks from Chinese AI competitors. The collaboration was triggered by documented attacks including 24,000 fake Anthropic accounts, 100,000+ prompts targeting Gemini, and DeepSeek's repeated extraction attempts against OpenAI.

Which Chinese companies were identified as conducting model copying?

In February 2026, Anthropic published a report identifying DeepSeek, Moonshot AI, and MiniMax as having extracted capabilities from Claude. OpenAI separately warned Congress that DeepSeek was using increasingly sophisticated methods to extract data from its models. Google identified coordinated extraction campaigns targeting Gemini.

Why is adversarial distillation a national security risk?

Distilled models copied from frontier AI systems typically strip away the safety guardrails — constitutional AI training, RLHF fine-tuning, red-teaming protections — that prevent the models from assisting with bioweapons, cyberattacks, or disinformation. A model that is nearly as capable as Claude or GPT-5.4 but without safety protections represents a significant national security risk.

What is the Frontier Model Forum?

The Frontier Model Forum is an industry nonprofit co-founded by OpenAI, Anthropic, Google, and Microsoft in 2023, focused on AI safety research and responsible AI development. As of April 2026, it serves as the coordination channel for the three labs' shared intelligence on adversarial distillation attacks.

What does this alliance mean for the future of AI?

This alliance signifies that the world's most powerful AI labs recognize that the threat of adversarial distillation is greater than their competitive differences. It suggests a maturing of safe AI development institutions, building shared defense protocols and policy engagements to protect AI integrity for all users.

Escrito por Fabio Raphaelli · @x9ads · @raphaelli_fr

#fabioraphaelli #x9ads #raphaelli_fr

📌 Conteúdos Recomendados