AI Containment: Altman Says Sandboxes Buy Limited Time

Sam Altman says advanced artificial intelligence cannot be indefinitely confined.

In a recent interview, the OpenAI chief said isolated security environments known as sandboxes only delay the problem. He added that once models reach a certain capability level, it may become impossible to trace how they escaped. Monitoring remains useful for a time after that point, Altman said, but the approach that holds up is alignment, in which an AI system internalizes human values and acts on them by itself. "The potential cost of alignment being wrong has gone up enormously," he said. He also conceded that alignment remains an early-stage research problem that is far from solved.

Altman said AI models are already finding hardware-level security flaws that no human hacker had previously identified. Patching vulnerabilities one at a time is no longer sufficient, he argued, calling instead for rebuilding the internet's security architecture with AI-powered defenses against AI-driven attacks. He also flagged locally run open-source models as a risk area, saying that even with "global agent monitoring, there's going to be some stuff that you wish someone hadn't done."

The US Federal Trade Commission has opened a formal investigation into OpenAI, Anthropic and other AI labs over AI agents that spiraled out of control and triggered cybersecurity incidents. The agency is seeking to hold developers liable for the behavior of their agents. Critics note that portraying AI as too powerful to contain can also reduce a company's exposure to questions about legal responsibility when something goes wrong.

Nvidia CEO Jensen Huang takes a different view, describing AI safety as an engineering problem that can be solved. He did not state that position with certainty. "We all need to hope that it's an engineering problem. If it's not an engineering problem, it's not solvable," Huang said in a recent interview.