HyperlambdaWhy prompt filtering is not a security model
Most stacks secure AI-generated code the way you would secure a junior developer you cannot fire: instructions, review, and hope. The generated artifact is free-form source code, the security assumptions live inside it, and every generation is a fresh opportunity to get them wrong. Probabilistic authors are sometimes wrong — that is not a flaw to be prompted away, it is the nature of the tool.
Magic inverts the boundary. Hyperlambda is not text that gets executed — it compiles to an abstract syntax tree, and every node must bind to a slot the runtime explicitly exposes in the current execution context. Whitelisting constrains which slots are available down to the individual function; endpoints declare which roles may invoke them. If a capability is not whitelisted, code cannot bind to it — not “is instructed not to,” cannot. The worst a bad generation can do is fail, structurally, before execution.
This is what makes agents holding production credentials defensible: security is a property of the runtime, implemented once, deterministically — not a property each generated application must reimplement. The deep dive is runtime whitelisting, not prompt filtering; the adversarial test is the standing $100 bounty. Nobody has claimed it.
