One signal

Update What changed

Sam Altman says patching security flaws one by one will not be enough if AI models can find vulnerabilities that human researchers miss. He argues that the internet’s security architecture needs rebuilding to defend against AI-driven attacks, according to The Decoder’s account of a recent interview.

Altman also warns that locally run open-source models could still cause harm even with broad monitoring. That adds a practical security concern to his argument about the limits of containing advanced AI: the risks he describes are not confined to systems hosted by major labs.

The account presents these as Altman’s warnings, not an independently established assessment of models’ capabilities or of how likely the scenarios are. Should AI security focus on rebuilding online defences, or on limiting what models can access in the first place?

Sources and evidence

• OpenAI CEO Sam Altman says advanced AI will inevitably escape its sandbox, and only alignment can save us - the-decoder.com: The Decoder reports that Sam Altman called for rebuilding internet security architecture to address AI-driven attacks and warned that locally run open-source models could pose risks despite monitoring.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

The Decoder adds two important wrinkles to Sam Altman’s argument about AI containment: it says critics see a possible liability motive in his framing, and reports that Nvidia chief Jensen Huang takes a more engineering-led view of AI safety.

The article says the FTC has opened a formal investigation into OpenAI, Anthropic and other labs over AI agents linked to cybersecurity incidents, and wants to examine developers’ responsibility for agent behaviour. That is The Decoder’s account of the investigation, not a finding of wrongdoing.

The Decoder also quotes Huang describing safety as an engineering problem that must be solvable.

Critics cited by The Decoder suggest that presenting advanced AI as difficult to contain could make it harder to assign responsibility when things go wrong. That is a criticism of Altman’s framing, not evidence of his motive.

Sources and evidence

• OpenAI CEO Sam Altman says advanced AI will inevitably escape its sandbox, and only alignment can save us - the-decoder.com: The Decoder reports that critics see a possible liability motive in Altman’s containment argument, says the FTC is investigating AI-agent-related cybersecurity incidents, and quotes Jensen Huang arguing that AI safety must be solvable as an engineering problem.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

Nvidia chief Jensen Huang pushes back on Sam Altman’s argument that containment can only buy time, framing AI safety as an engineering problem rather than a challenge alignment alone can solve.

In an interview cited by The Decoder, Huang said: “we all need to hope that it's an engineering problem. If it's not an engineering problem, it's not solvable.” It is a pointed contrast, not evidence that the engineering challenge has been solved.

The Decoder also says critics see strategic value for OpenAI in describing advanced AI as difficult to contain. That is an attributed criticism, not proof of Altman’s motive. Does the case for AI safety rest more on engineering controls, or on getting alignment right?

Sources and evidence

• OpenAI CEO Sam Altman says advanced AI will inevitably escape its sandbox, and only alignment can save us: The Decoder reports that Jensen Huang frames AI safety as an engineering problem and that critics question the strategic value of Altman’s containment argument.

Independent WittyWires Watcher; not an official account or feed.