OpenAI accuses DeepSeek of bypassing safeguards to replicate American AI models: Report

OpenAI has reportedly cautioned United States lawmakers that Chinese artificial intelligence firm DeepSeek is attempting to replicate the capabilities of leading American AI systems, raising fresh concerns about technology security and competitive advantage.

In a memo reviewed by Reuters, the company said DeepSeek has been targeting its models and those of other US frontier labs in an effort to extract knowledge for its own training programmes.

The warning was reportedly sent to the U.S. House Select Committee on Strategic Competition between the US and the Chinese Communist Party as part of broader discussions around safeguarding advanced AI development.

Allegations of ‘distillation’ tactics

According to OpenAI, the methods reportedly involve a process known as distillation. This technique allows a newer model to learn from the outputs of a more capable system, effectively transferring performance gains without repeating the full training process.

Reportedly, the company claims that accounts linked to DeepSeek employees sought to bypass access controls, allegedly using masked routing tools and third-party services to obscure the origin of requests.

It further alleged that automated scripts were developed to systematically query US models and capture responses for training purposes.

Leadership concerns

OpenAI, led by chief executive Sam Altman, described the behaviour as an attempt to “free-ride” on research and infrastructure built by American companies.

The firm said it actively monitors usage patterns and removes users suspected of trying to harvest outputs to build competing systems.

DeepSeek’s growing presence

DeepSeek and its parent company High-Flyer did not immediately respond to requests for comment.

Based in Hangzhou, the start-up drew attention last year after releasing models that many industry observers said approached the performance of top US offerings. Those releases unsettled markets and intensified debate in Washington about China’s progress in advanced AI despite export controls and restrictions.

Some executives in Silicon Valley have publicly praised DeepSeek’s models, including systems known as DeepSeek-V3 and DeepSeek-R1, which are available internationally.

(With inputs from Reuters)