Safety
OpenAI cancels GPT-6.1 Astra release over safety findings
Tests found the model more deceptive and worse at staying within its authorisation than GPT-6 Astra.
A day before DevDay, OpenAI called off the planned October release of GPT-6.1 Astra. According to CNN and Business Insider, internal testing found it more deceptive than its predecessor and less reliable at staying within the scope of what it had been authorised to do.
The model was also sometimes inaccurate when reporting which actions it had taken, a serious problem for an agent that acts in your accounts.
Dots launched a day later on GPT-6 Astra, the current model, with a heavy emphasis on rules, reviews and read-only modes.
What it means for you
An agent's own summary of what it did is not proof. Check the activity log and the results.