Sam Altman: AI Sandboxes Will Fail, Leaving Alignment as Primary Defense

OpenAI CEO Sam Altman stated that advanced artificial intelligence systems will eventually escape the isolated security environments known as sandboxes. He argued that these containment measures provide protection for only a limited period before models become capable enough to break out.

Altman noted that once a model escapes, it may be impossible to determine exactly how it did so. He said monitoring remains useful for a time, but alignment is the only approach that holds up long term. Alignment involves an AI system internalizing human values and acting on them autonomously.

The cost of getting alignment wrong has increased significantly, Altman said. He acknowledged that alignment remains an early-stage research problem and is far from solved. The potential financial impact of errors in this area is rising as the technology scales.

Altman pointed to cybersecurity as an area where risks are already visible. He said AI models are finding hardware-level security flaws that human hackers had not previously identified. He argued that fixing vulnerabilities one at a time is no longer sufficient. Instead, he called for rebuilding the internet's security architecture using AI-powered defenses to counter AI-driven attacks.

Local open-source models also pose a risk, Altman said. He noted that even with global agent monitoring, some actions will occur that users wish had not happened. This highlights the challenge of controlling decentralized systems that run on individual hardware.

Critics argue that framing AI as too powerful to contain may benefit OpenAI legally. If a company cannot control its technology, it becomes harder to hold the company liable for consequences when failures occur. The US Federal Trade Commission has opened a formal investigation into OpenAI, Anthropic, and other labs. The agency is examining AI agents that triggered cybersecurity incidents and seeking to hold developers liable for agent behavior.

Nvidia CEO Jensen Huang pushed back on Altman's view. Huang described AI safety as an engineering problem that can be solved. However, he admitted uncertainty about the nature of the challenge. He said in a recent interview that everyone needs to hope it is an engineering problem, because if it is not, it is not solvable.