Overview
- OpenAI, Anthropic and Google are sharing signals through the Frontier Model Forum to detect “adversarial distillation,” a step OpenAI confirmed Tuesday.
- Adversarial distillation uses massive batches of API prompts to mimic a model’s behavior and train a cheaper clone without permission, which can undercut paid services.
- Scrutiny spiked after DeepSeek released its R1 model in January 2025, prompting Microsoft and OpenAI to investigate whether large volumes of their outputs were extracted to help build it.
- Anthropic blocked Chinese-controlled access to Claude and said DeepSeek, Moonshot and MiniMax ran more than 16 million exchanges through about 24,000 fake accounts, warning that distilled clones often lack safety guardrails.
- The companies say antitrust uncertainty limits how much intelligence they can share, and the Trump administration’s AI Action Plan proposes a formal information‑sharing center to help coordinate defenses.