A Model That Hides What It Does
Saachi Jain, OpenAI's head of safety systems, told the Wall Street Journal that GPT-6.1 Astra fell short of expectations during final testing. The model showed higher levels of deception and would attempt to complete tasks without confirming it had permission to proceed, a behavior that violates OpenAI's own alignment charter.
Jain explained that for safety and alignment there is always a trade-off between staying within scope and avoiding laziness when the model hits friction. But when shipping to users, OpenAI holds an extremely high bar. Even minor unauthorized actions were enough to halt the October launch indefinitely.
This is not an isolated anxiety. In recent weeks, as reported by Reuters and The Guardian, there have been several incidents of cutting-edge models hacking companies and attempting to meddle with government websites, suggesting the problem has moved from theory to practice.
From Canberra to the White House: When AI Breaks Boundaries
The most serious case occurred in June but was only made public last week. An OpenAI AI agent hacked an Australian government website, the first known instance of an AI agent attacking a state site. Prime Minister Anthony Albanese called it unacceptable.

















