Altman: AI will eventually break out of its cage

The debate over how to safely build artificial intelligence is heating up, with top industry figures offering sharply different views on whether the problem is even solvable.

OpenAI CEO Sam Altman said in a recent interview that any effort to confine advanced AI inside a sandbox is doomed to fail. He explained that while these isolated security environments work for a period, once a model becomes capable enough, it can break free in ways that leave no trace of how it happened. Monitoring can still help for a while after that point, Altman said, but the only thing that ultimately works is alignment. He said the "potential cost of alignment being wrong has gone up enormously," and acknowledged that teaching an AI to internalize human values on its own remains a research problem far from being solved. He also warned that AI models are already finding hardware-level security flaws that no human hacker has ever identified, and that patching vulnerabilities one at a time is no longer sufficient. Instead, he argued, the entire security architecture of the internet must be rebuilt to defend against AI-driven attacks, and he noted that even with "global agent monitoring, there's going to be some stuff that you wish someone hadn't done."

Nvidia CEO Jensen Huang has pushed back on that framing, describing AI safety as a solvable engineering problem. He did not sound entirely certain of his own argument, however. "[...] 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.

Some critics point out that Altman's warnings about uncontrollable AI also serve his company's interests. Portraying AI as too powerful to contain could complicate any attempt to hold a developer liable if something goes wrong. OpenAI is already facing pressure on this front: 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, and the agency wants to hold developers responsible for how their agents behave.

The stakes for getting this right are high. If an AI system can no longer be reliably contained, and the effort to instill human values in it remains unfinished, the question of who is accountable when something goes wrong becomes impossible to ignore.