Distillation attacks involve querying a model extensively to extract its outputs and behavior, then using that data to train a competing model at a fraction of the original cost. The agencies claim the Chinese companies leveraged this technique to effectively replicate the capabilities of leading U.S. models, potentially undermining the competitive edge of American AI developers. The extracted data could also be used to map model weaknesses or bypass safety guardrails.

While the official statement does not name the specific models targeted, the scale of token extraction points to large-scale, automated abuse of API access. The advisory emphasizes that such attacks are not merely intellectual property concerns but also pose systemic risks to AI supply chains and national security. Affected organizations are urged to monitor for anomalous API usage patterns, such as high-volume requests from single sources, repeated identical prompts, or sudden spikes in token consumption.

• Monitor API logs for traffic originating from known hosting ranges or VPN endpoints.
• Implement rate limiting and anomaly detection on model endpoints to flag batch extraction attempts.
• Review access logs for patterns consistent with systematic prompt repetition.
• Enforce stricter authentication and usage policies for high-privilege API keys.

The advisory reinforces a growing concern among security researchers that model distillation is evolving from a niche technical issue into a staple of state-sponsored economic espionage. This is not the first time U.S. authorities have warned about AI-related IP theft, but the explicit mention of billions of tokens marks a significant escalation in the reported scope of such operations.

Source: Unknown

Does your organization have visibility into API-level token usage, and what thresholds would you set to detect a distillation attempt before it reaches billions of requests?