Science & Tech
Google Limits New AI Model To Cyber Defenders
Google is testing safeguards before wider access to developers and consumers.
Google is testing safeguards before wider access to developers and consumers.
Google is holding back its newest from the public because it worries the tool could be used for and other harmful tasks. Instead of a full launch, the company is letting a small group of trusted security experts test it first. Google is also checking its safety limits and getting feedback from the and its own teams.
The choice shows how major AI releases are slowing as the risks grow. Google says the model is built to reject harmful requests, but critics warn that strong tools can still help people attack banks, hospitals, or public systems. For everyday people, the trade-off is clear. Faster AI can improve defenses, but if the wrong people get it too soon, it can also make damage worse.
Google said it would widen access only after further testing and were complete. The company framed the launch as a that had to be slowed by safety review, not treated like a standard product release.
Google said it was participating voluntarily in the U.S. government’s . The company said that government feedback would be part of the broader safety review before Argon reaches developers, enterprises, and consumers.
Google said the first rollout was going to and internal teams, so the model could be tested in controlled conditions before wider access. Coverage of the launch said the company would use feedback from these early testers to refine and expand access later.
Google said Argon was designed to refuse requests tied to cyberattacks and to chemical, biological, radiological, and nuclear, or , weapons. The company also said it was monitoring reasoning and , and that it had built the release around a .
Google said the restriction was meant to reduce the chance that the model could be . Reporting on the launch said the company and outside experts were worried the system could help target .
Google announced , but did not release it to the general public. Instead, the model was opened first to a of cybersecurity experts through its , with Google saying the rollout would be phased.