Read full article
strict (default): Rejects writes when the buffer is full and too many writes are pending. Catches "fire-and-forget" patterns where producers ignore backpressure.
，详情可参考51吃瓜
But Anthropic also imposed limits that Michael views as fundamentally incompatible with war-fighting. The company’s internal “Claude Constitution” and contract terms prohibit the model’s use in, for instance, mass surveillance of Americans or fully autonomous lethal systems—even for government customers. When Michael and other officials sought to renegotiate those terms as part of a roughly $200 million defense deal, they insisted Claude be available for “all lawful purposes.” Michael framed the demand bluntly: “You can’t have an AI company sell AI to the Department of War and [not] let it do Department of War things.”
港交所2025年净赚177.5亿港元
The common pattern across all of these seems to be filesystem and network ACLs enforced by the OS, not a separate kernel or hardware boundary. A determined attacker who already has code execution on your machine could potentially bypass Seatbelt or Landlock restrictions through privilege escalation. But that is not the threat model. The threat is an AI agent that is mostly helpful but occasionally careless or confused, and you want guardrails that catch the common failure modes - reading credentials it should not see, making network calls it should not make, writing to paths outside the project.