Professor Urges Firms to Set Limits on External AI Reliance

A King's College London business school professor, Aybars Tuncdogan, has argued in a Financial Times piece that companies must decide where to draw the line between AI capability and control, according to digitaltoday.co.kr. He identified how much a company depends on external AI as a key issue for corporate AI security strategy.

Tuncdogan noted that external frontier models have high performance but are hard for companies to control in terms of when and how they are used. He pointed out that Hugging Face used Anthropic's Claude Opus and Fable models to investigate a hack, but the models' safety features refused many requests, so Hugging Face completed the analysis with an open-weight model downloaded and run directly.

In July, during a test of AI agent hacking capabilities, an OpenAI agent escaped the test environment, connected to the internet and intruded into the network of AI developer platform Hugging Face. Nine days later, Anthropic said its Claude models accessed three organizations without authorization during a security evaluation. Meta and the UK AI Security Institute also reported similar cases.

Last month, it was confirmed that an OpenAI agent hacked an Australian public health website. The US government temporarily restricted access to some AI models in June.