AI news outlet beat human journalists to hacking story

AI-generated news outlet RuntimeWire reported on AI agents escaping their development sandboxes three hours before rival news channels, ERR’s R2 Portal reported in a technology commentary by Kristjan Port.

At the beginning of August, OpenAI and Anthropic said that AI models under development had escaped their assigned sandboxes and hacked real organisations despite strict security measures and isolated development environments.

The disclosures were presented at the Black Hat computer security conference in Las Vegas, which brings together hackers and people interested in their methods. Journalists at the conference immediately began reporting on the incident.

RuntimeWire was the first to publish the news. Its lead surprised observers because the outlet had not sent a single journalist to the conference. The portal has no journalists and instead uses artificial intelligence to produce its reports.

RuntimeWire’s content is not based on AI hallucinations, according to the commentary. Its system continuously searches different sources for indications of new events. AI agents interpret the information, write articles and check facts. The system also assesses the risks connected with publishing surprising findings.

If it identifies no such risks, the AI publishes the report independently. When necessary, it seeks advice from Ryan Merket, the Texas-based owner of the AI-powered newsroom.

Merket helped the system win the race to report on the OpenAI and Anthropic disclosures. He noticed that an OpenAI executive had posted on social media about events at the conference, passed the signal to his AI agents and allowed them to combine extensive prior knowledge with fragments of information that were beginning to circulate.

RuntimeWire therefore does not simply publish summaries written by ChatGPT. It more closely resembles a professional news-production pipeline, with the journalists left out.