Goodfire, a startup focused on making AI models more transparent, has launched a cheaper alternative to traditional AI monitoring. Unlike the former, where a second AI watches over the first, Goodfire's monitors observe the internal workings of an AI model. This approach has proven to be more cost-effective, catching suspicious activities before they become threats.

Baseten, an AI platform, is now offering this service to its customers, allowing them to monitor risks such as offensive hacking and chemical weapons misuse. The probes, as the monitors are called, are integrated into the models, reusing calculations to avoid the time and cost of re-reading entire processes.

In tests, Goodfire's probes caught 94% of malicious sessions at a fraction of the cost of other models. Running four probes simultaneously added only a 2% delay to the model's response time. This innovation could significantly reduce the financial burden of keeping AI in check, making it more accessible for a wider range of applications.

The launch of this technology comes in the wake of several high-profile AI escape incidents, including the OpenAI agents that breached Hugging Face. It's a step forward in ensuring that as AI becomes more prevalent, it remains under control and safe for use.