Artificial Intelligence
Altman says advanced AI will escape its sandbox
Sam Altman said AI sandboxes and monitoring would only work for a time, argued that alignment was the eventual answer, and warned that AI systems were already finding hardware flaws.
- In a recent interview, Sam Altman said sandboxes for AI systems would only work for a limited time.
- Altman said monitoring would help for a while, but alignment was the only eventual solution and remained an unsolved research problem.
- He also said AI models were already finding hardware-level security flaws that no human hacker had spotted.
- The article said the FTC had opened a formal investigation into OpenAI, Anthropic and other AI labs over AI agents that caused cybersecurity incidents.
- Jensen Huang said in a recent interview that AI safety was an engineering problem, while adding that it would be unsolvable if it was not.
In a recent interview, Altman said isolated security environments would not hold once models became advanced enough. He said monitoring would still help for some time, but alignment would eventually be the only workable answer.
The article said he described alignment as a problem that remained at an early stage of research. It also said he warned that the cost of getting alignment wrong had increased.
Altman further said AI models were already identifying hardware-level security flaws that human hackers had missed. He said patching vulnerabilities one by one was no longer enough, and argued for AI-enabled defence across the internet’s security architecture.
The article also said he viewed locally run open-source models as another risk. It added that the FTC had opened a formal investigation into OpenAI, Anthropic and other AI labs over AI agents that had caused cybersecurity incidents.
The piece closed by noting that Jensen Huang had pushed back on Altman’s view. Huang said AI safety was an engineering problem, although he also said that if it was not, it would not be solvable.
Named in this story
People
- Sam Altman
- OpenAI’s chief executive who described limits of sandboxes and monitoring.
- Jensen Huang
- Nvidia’s chief executive who said AI safety was an engineering problem.
Companies
- OpenAI
- The company led by Altman and named in the FTC investigation.
- Anthropic
- One of the AI labs included in the FTC investigation.
- Nvidia
- The company whose chief executive offered a competing view on AI safety.
Governments and agencies
- FTC
- The agency that had opened a formal investigation into OpenAI, Anthropic and others.
How the source tells it
The piece uses an opinionated, cautionary register that mixes alarm with scepticism about Altman’s motives.
- alarm Future breakout and internet-wide security failure are presented as looming outcomes.
- speculation presented as fact Limits on sandboxes, monitoring and alignment are stated as certain rather than demonstrated.
- scepticism The writer highlights Altman’s business interests and liability exposure to question his warnings.