Sam Altman 称先进 AI 终将逃出沙箱，只能靠对齐兜底

OpenAI CEO Sam Altman believes advanced AI can't be kept locked up. He has strategic reasons to frame AI this way, but his warnings about what these models can do deserve attention.

In a recent interview, Altman explains that sandboxes, the isolated security environments used to contain AI systems, only work for a limited time. Once models get smart enough, there's no way to even trace how they broke out.

Monitoring still helps for a while after that, he says. But eventually, the only thing that works is alignment, meaning the AI has to internalize human values on its own and act accordingly. Altman says that "the potential cost of alignment being wrong has gone up enormously," though that remains an early-stage research problem, far from being solved.

Altman also warns that AI models are already finding hardware-level security flaws no human hacker ever thought of. Patching vulnerabilities one by one isn't enough anymore, he says. Instead, the entire security architecture of the internet needs to be rebuilt, with AI-powered defense against AI-driven attacks. He sees locally run open-source models as a risk too, noting that even with "global agent monitoring, there's going to be some stuff that you wish someone hadn't done."

Altman's warnings also serve his business interests

Critics will point out that Altman's talk about uncontrollable AI benefits his own company. Framing your AI as too powerful to contain could be a way to preempt liability questions. If even the company that built the AI supposedly can't control it, it gets a lot harder to argue that the company should be held responsible when things go wrong.

OpenAI is under real pressure here. The FTC has opened a formal investigation into OpenAI, Anthropic, and other AI labs over AI agents that spiraled out of control and caused a wave of cybersecurity incidents. The agency wants to hold AI developers liable for how their agents behave.

Altman is getting some pushback from Nvidia CEO Jensen Huang, who frames AI safety as a solvable engineering problem. Huang didn't sound entirely convinced by his own argument, though. "[…] 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.

AI News Without the Hype – Curated by Humans

Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.