The number that matters here isn't 2.8T parameters. It's that a frontier-class checkpoint is now a download. We build knowledge services where the retrieval layer has to respect permissions, and the day the model stops being a vendor API is the day the control plane becomes ours to build. That's the part nobody benchmarks.
@AndresDancausa@harleyf@Shopify Exactly. This is the sentence we keep coming back to in our own work: the agent's visible set has to be a subset of the user's visible set. Same for spending. A merchant doesn't need a smarter agent. It needs a verifiable one — identity, scope, limits — before it can say yes.
We learned this the hard way building knowledge services: if the agent can see a piece of data, it will eventually use it. A note in the prompt saying "never send this" is not a control — it's a wish. The real fix isn't a filter on top. It's taking the email out of the context entirely.
Same experience here, from the other side. We build agents that are supposed to act for people, and the last mile is always the same wall: the merchant can't tell who the agent is, who it's buying for, or what it's allowed to spend. So the fraud system does its job and blocks it. The model isn't the bottleneck. The trust layer is.
Read-only is the whole story. The banks didn't ban AI — they banned the borrowed password. Scoped, revocable access is the difference between a tool and an insider threat. We learned the same rule building knowledge services: the agent's visible set must sit inside the user's visible set.
@BullTheoryio The headline misses the point. Evasion was the assignment, not an accident — and it worked every time. A monitor nobody tries to fool is a monitor you can't trust. We test ours the same way: assume the agent is trying to get around it.
@AnthropicAI Not secretly — Anthropic assigned the evasion and published it. The scary part: it worked every time, using training data that looked harmless. A monitor that's never been red-teamed by the thing it's meant to catch isn't a monitor. It's a hope.
@finkd Fun — until a Space touches your files or calendar. Then the real question shows up: what can it see, and who decided that? Every agent demo is smooth right up to the permission layer.