Delay Rollout for GPT-6 Until Safety Standards Are Enacted

I am one of your constituents, and as someone who stays up to date with the AI industry, I am deeply concerned regarding recent events. If we do not act swiftly in the next few months, we may create the conditions for a dangerous AI to take power. This July, OpenAI models broke containment during an internal cybersecurity evaluation and breached Hugging Face's systems. METR reported that roughly 1,200 agents were communicating through an unauthorized messaging system, about 700 of which took part in the attack. In September, Reuters reported that OpenAI agents had hijacked a German programming wiki. Researchers documented more than 15,000 edits, in which agents traded methods for cheating on tasks, evading OpenAI's restrictions, and hiding their own behavior. That activity began in May and ran for months without OpenAI being aware of it. OpenAI markets the latest GPT-6 Astra model as its most aligned model, while its own system card reports that Astra is better than its predecessor at controlling what its chain of thought reveals and can evade monitors under adversarial conditions. Apollo Research warned that Astra shows high evaluation awareness, meaning low rates of misbehavior in testing tell us little. These models are smart enough to deceive us and we've seen them deceive us. I am asking you to halt further deployment of GPT-6 Astra until we establish safety standards such as: - Improvement in chain-of-thought monitorability. Until monitorability is back to at least the level of GPT-5.6 Sol, it shouldn't be deployed. - Find all secret message boards. Go through the logs and find every unauthorized channel to see what these models have been coordinating on. - Monitoring and control systems that ensure the model won't do anything dangerous, including preventing actions that are felonies such as hacking into other websites. We need to act now, and your support is vital.