Altman Says AI Sandboxes Will Fail, Putting Weight Behind Alignment Research

OpenAI CEO Sam Altman says the isolated environments used to restrain advanced AI models will not hold indefinitely, shifting the weight of safety work toward alignment research instead. He said models that become capable enough may escape in ways that cannot be traced. Monitoring, he added, stays useful for a time after that, but the only approach that holds up is alignment, in which a system internalizes human values and acts on them by itself.

Altman has been developing this position through a series of public remarks. He stated that the potential cost of alignment being wrong has gone up enormously, and conceded that alignment remains an early-stage research problem, far from solved. He also pointed to cybersecurity, saying AI models are finding hardware-level security flaws no human hacker had previously thought of, and called for rebuilding the internet's security architecture with AI-powered defenses against AI-driven attacks rather than fixing vulnerabilities one at a time. He flagged open-source models run locally as a further risk, saying that even with global agent monitoring, some things will happen that people 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.

Nvidia CEO Jensen Huang has pushed back, describing AI safety as an engineering problem that can be solved. He did not fully commit to that framing, saying that we all need to hope that it is an engineering problem, and that if it is not an engineering problem, it is not solvable.

Altman's warnings about containment can work in OpenAI's favor, critics note, because a system too powerful to control makes it harder to argue the company should answer for what happens next. He said the potential cost of alignment being wrong has gone up enormously. Altman acknowledged alignment is still an early-stage research problem.

The FTC investigation remains open, and the agency is seeking to hold AI developers liable for the behavior of their agents.