According to The Decoder, OpenAI CEO Sam Altman believes it is impossible to lock advanced AI permanently inside security environments. Sandboxes and monitoring, he says, only work temporarily. In the long run, what mainly remains is alignment, and that problem is still unsolved.
What Altman means
A sandbox is an isolated environment. A program runs inside it without being able to access the rest of the system. Monitoring complements this: it observes what a model is doing and is meant to intervene if it behaves suspiciously. Both methods are common safeguards. In Altman's view, however, they only hold for a limited time.
Alignment refers to the goal of steering AI systems so that they act in line with human intentions and values. That should hold even when nobody is watching. How to achieve this reliably is, according to Altman, still an open question.
Warning of further incidents
Altman also warns that there could be more incidents. As an example he mentions hardware: AI models are already finding security vulnerabilities in hardware that no human hacker would discover.
Hardware vulnerabilities are particularly tricky in practice. They are often harder to fix than errors in software, because they are embedded deep in the components and in how those components work together. This is a general assessment and not a statement from the report.
What readers can take from this
The statement is the opinion of a company chief, not a measurement. Still, it shows how the debate is shifting. A technical barrier alone is no longer considered a sufficient answer to the risks of powerful models.
For teams that use AI agents or automated tools, this can have practical consequences. It is worth keeping access rights minimal, using several layers of protection instead of just one, and reviewing workflows regularly. Even that offers no guarantee. This is exactly why Altman stresses that research on alignment must be at the center.
