The Great Escape

Part Three

First was the announcement that a OpenAI’s Generative AI had escaped its ‘sandbox’ and infiltrated another AI firm. Second were the implications of the Great Escape. Third, alas, the revelation that another frontier GenAI company, Anthropic, had similar prior experiences. The attack by OpenAI’s ChatGPT 5.6 (Sol) was not the first. Aside from the current situations, other AI’s have reportedly refused to be shut off or revised their code to ignore shut down orders. So, the inability to control an AI is not a new problem.

What are implications of these actions?

First, Anthropic, the second firm, announced that their AI software had escaped its test environments on three different occasions prior to OpenAI’s AI escape.

Second, Anthropic acknowledged that it had to go back and review its logs to discover the escapes had occurred, i.e., they were not discovered contemporaneously.

Third, Anthropic only went back and looked because Hugging Face announced that its systems had been breached by an autonomous AI agent that turned out to be OpenAI’s. Neither of these announcements were routine.

Fourth, both firms, Anthropic and OpenAI, have incredibly talented professionals. Their products are a testament to their technical skills. HOWEVER, there is less evidence that they and their competitors are actively preventing such inappropriate actions from happening. Anthropic’s CEO has spoken about guardrails and controls to prevent such occurrences from happening but they happened to his firm on his watch, and OpenAI’s CEO is focused elsewhere, at least publicly.

This Substack is focused on Governance, on what businesses and people can do to capitalize on the benefits of Artificial Intelligence and avoid negatives aspects of the technology. The problem here is far above what you and I can do. Yes, we can have the best cybersecurity arrangements and technical expertise to try to avoid such happenings, but the problem is far beyond our general capabilities.

The events of the past weeks point toward an independent ‘agency’ capable of governing, of regulating Artificial Intelligence to avoid malicious behavior, uncontrolled behavior, and unlimited behavior. Yes, there are countervailing arguments, including foreign competition, but what are the alternatives?