What containment can and cannot do
In a recent interview, Altman said AI sandboxes—isolated environments intended to contain systems—work only for a limited time. Once models become sufficiently capable, he argued, people may no longer be able to track how they got out.
Monitoring might help for a while. Altman’s proposed long-term answer is alignment: AI systems would need to internalize human values and act accordingly. He said the cost of an alignment failure has risen sharply, even though the field remains at an early stage and is far from solving the problem.
Altman also warned that AI models are finding hardware vulnerabilities that no human hacker had thought of. Fixing flaws one by one is not enough, in his view. He called for rebuilding internet security architecture and using AI to defend against AI-powered attacks.
He sees risks in locally run open-source models, too. Even with global monitoring of AI agents, Altman said, some actions will happen that people later regret.
The accountability question
Critics argue that warnings about uncontrollable AI serve OpenAI’s interests. If a company presents its systems as too powerful to contain, that can weaken scrutiny of its responsibility: if even the developer cannot control the technology, it may be harder to hold the developer accountable for what follows.
That argument matters now because OpenAI is under pressure. The US Federal Trade Commission has opened an official investigation into OpenAI, Anthropic and other AI labs. It concerns agents that went out of control and triggered a wave of cybersecurity incidents. The commission intends to hold AI developers responsible for their agents’ behavior.
Nvidia CEO Jensen Huang has pushed back on Altman’s framing, saying AI safety is a solvable engineering problem. But in a recent interview, Huang did not sound entirely certain. He said everyone has to hope the problem really is an engineering one; if it is not, it cannot be solved.
I think the tension is less about whether safeguards can fail than about what follows from that possibility. If alignment is the only lasting defense, as Altman argues, responsibility does not disappear when containment does. The more important test is whether companies accept accountability for systems they say may eventually outrun their control.
Source: the-decoder.com
Daily AI news
Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.
Only what matters — every day
Follow on X