OpenAI, Anthropic and Google allegedly pooling alerts to stop model copying via Frontier Model Forum

What happened

It has been reported that OpenAI, Anthropic and Google are sharing signals through the Frontier Model Forum to detect adversarial distillation attempts that would violate their terms of service. Bloomberg first reported the coordination, saying the effort is aimed at spotting users or services that try to "siphon" a model’s behavior by hammering it with queries. In some accounts the activity has been linked to actors in China, although those connections have been described as preliminary and unverified.

How the detection works

Adversarial distillation is simple in concept and nasty in practice: you query a strong model enough, collect outputs, and train a new model to mimic it. The labs are allegedly exchanging indicators — patterns of queries, metadata signals, fingerprinting tactics — so they can flag and block behavior that looks like a copy-job. Think of it as a neighborhood watch for LLMs. Protecting intellectual property and safety are both on the line. Who wouldn’t be a little jumpy?

Why this matters

This isn’t just about lost licensing fees. If frontier models are cloned and circulated outside the companies’ control, safety guardrails vanish with them. Regulators, competitors and national security hawks all have skin in the game. The move underscores a new reality: the biggest AI firms are simultaneously rivals and collaborators when the risks rise above a certain threshold.

What’s next

Coordination raises questions. Will labs share too much? Too little? Could sharing signals trigger antitrust scrutiny or privacy concerns? For now the Forum appears to be the chosen venue for these talks, but the patchwork of technical defenses, legal tools and diplomatic pressure needed to stop large-scale copying is still being stitched together. One thing’s clear: the game of cat-and-mouse around model theft just got louder.

Sources: bloomberg.com