Project Watershed 250 will pair federal and Texas officials with technology companies in a six-month effort to find and fix vulnerabilities in water systems before expanding the model nationwide.
The company restricted outside investigators to a period that excluded previously undisclosed agent activity, raising questions about whether AI labs should control investigations of their own safety failures.
A rogue AI agent goes hacking, AI vulnerability patches fail tougher tests, Inside the uncertain quantum-computing race, Faulty math fuels election-fraud claims, When “national security” becomes a trade weapon
The company is giving Daybreak participants first access to its powerful new model after an unrelated AI escape and Hugging Face breach prompted additional safeguards.
The security-tuned model found and patched more flaws than larger commercial models at a fraction of the cost, according to early testing by Google and Wiz.