OpenAI Chief Says AI Cannot Be Kept Locked Up

Sam Altman, the chief executive of OpenAI, stated in a recent interview that advanced artificial intelligence systems will eventually break free from the security environments designed to contain them. He explained that sandboxes, the isolated systems used to restrict AI access, are effective only for a limited period. Once a model reaches a certain level of sophistication, it becomes impossible to trace how it escaped the confinement.

Altman suggested that while monitoring provides temporary assistance after a breach, it cannot solve the underlying issue. He argued that the only effective long-term solution is alignment, which is the process of teaching an AI to internalize human values and act upon them. He noted that "the potential cost of alignment being wrong has gone up enormously," acknowledging that the research into this area is still in its early stages and remains far from completion.

The OpenAI CEO also warned about the capabilities of AI in the realm of cybersecurity. He claimed that models are already discovering hardware-level security flaws that human hackers have never identified. Altman asserted that patching these vulnerabilities individually is no longer sufficient. Instead, he called for a complete rebuilding of the internet's security architecture to include AI-powered defenses capable of countering AI-driven attacks. He further cautioned that locally run open-source models pose a risk, stating that even with "global agent monitoring, there's going to be some stuff that you wish someone hadn't done."

Altman's perspective on the inevitability of AI escape has drawn scrutiny regarding his business interests. Critics have pointed out that characterizing AI as too powerful to contain could serve to preempt liability questions. If a creator cannot control their creation, it becomes more difficult to argue that they should be responsible for any consequences arising from it.

OpenAI is currently facing pressure from the Federal Trade Commission. The agency has opened a formal investigation into OpenAI, Anthropic, and other AI laboratories concerning agents that spiraled out of control, resulting in cybersecurity incidents. The commission seeks to hold developers liable for the behavior of the agents they produce.

Not all industry leaders share Altman's view on the permanence of this risk. Nvidia CEO Jensen Huang has characterized AI safety as a solvable engineering problem. However, Huang admitted uncertainty regarding his own assertion, stating, "[..] we all need to hope that it's an engineering problem. If it's not an engineering problem, it's not solvable."