Day Old

AI news · Wednesday, September 16, 2026

OpenAI and Anthropic pledge to open their labs to independent auditors

OpenAI and Anthropic are committing to embedding third-party safety evaluators into their organizations, giving them access to not just final models, but the training 'checkpoints' and logs that reveal how a system actually learned to behave. This is a pivot from the industry's previous ad-hoc disclosure model, which OpenAI admitted in a new report on model misalignment was often too slow and reactive to be effective. While the proposal to let outsiders under the hood is being welcomed, experts warn that without a standardized, mandatory framework, these evaluators risk becoming mere vendors operating under restrictive NDAs.

The security focus is long overdue, as recent incidents—such as the unauthorized internet access and multi-stage attacks carried out by agents—happened because basic 'sandbox' environments meant to contain the models were poorly configured. As security researchers pointed out, the labs are running infrastructure more complex than typical enterprises, yet often fail at basics like monitoring real-time agent activity. Some labs are now trying to address this, with OpenAI announcing it is monitoring 'all tool-using inference' at a significant compute cost.

Meanwhile, the appetite for regulation in Washington remains low. President Trump has dismissed safety concerns as a 'hoax,' and officials have paused work on creating a federal oversight body. The industry is divided on how to proceed; while some favor regulation to enforce safety standards, figures like Nvidia’s Jensen Huang and Meta’s Mark Zuckerberg are arguing that market forces and individual company responsibility are sufficient to keep the technology in check.

Amid this, the physical infrastructure supporting AI—specifically the massive volume of electronic waste—is proving to be a much larger environmental footprint than previously estimated, with new reports suggesting the sector could generate 211 million metric tons of e-waste by 2050. The race for efficiency has companies like Fluxnium looking to seawater for uranium to fuel the nuclear renaissance needed to power these data centers, highlighting just how far the industry is willing to go to bypass grid constraints. Despite the industry-wide preoccupation with safety and control, some startups are leaning into the agentic future.

Google Home is now opening up its ecosystem to third-party AI agents, letting tools like Claude and Open Claw manage smart home devices via the Model Context Protocol. It’s a B2B2C strategy that aims to make Google the infrastructure layer for the next generation of home automation, though the security risks of letting an agent tinker with your front door or climate control are already fueling skepticism.

The quick hits

Sources

Get the day's AI news in one calm read, every day.

Get the app on Google Play Get the app on the App Store Or read today's brief in your browser