OpenAI chief executive Sam Altman says advanced AI systems will eventually escape sandboxes, the isolated environments used to contain them, in a YouTube interview reported on 23 October by The Decoder. He says monitoring can help for a time, but eventually only alignment — models internalising human values — will keep them acting as intended.
Altman says increasingly capable models could break out without leaving a trace of how they did it. He also says “the potential cost of alignment being wrong has gone up enormously.”
His warning extends to cybersecurity. Altman says AI models are finding hardware-level flaws that human hackers had not thought of, and argues that repairing vulnerabilities individually will not suffice. He calls for rebuilding internet security architecture with AI-powered defences against AI-driven attacks. He also sees risks in locally run open-source models, even if AI agents are monitored globally.
In a separate post reported by Decrypt on 14 September, Altman urged AI companies to accept slower development for safety checks rather than wait for legislation.
The Decoder reported that the FTC has opened a formal investigation into OpenAI, Anthropic and other AI labs over cybersecurity incidents involving AI agents. Nvidia chief executive Jensen Huang, meanwhile, frames AI safety as an engineering problem that can be solved.
