import { AlertTriangle, KeyRound, FolderOpen, FileCode } from 'lucide-react' const exposures = [ { icon: KeyRound, path: '~/.ssh/', label: 'SSH keys', desc: 'Private keys, known hosts, agent configs', }, { icon: FileCode, path: '.env', label: 'Environment secrets', desc: 'API keys, database URLs, auth tokens', }, { icon: FolderOpen, path: '~/*', label: 'Full filesystem', desc: 'Every repo, document, and config file', }, ] export function Problem() { return (
The problem

Every agent inherits everything.

AI coding agents run as your user. They see your SSH keys, cloud tokens, env files, and entire home directory. The model decides what to access at runtime, guided by weights you didn't train, at machine speed. One wrong inference is all it takes.

{exposures.map((item) => (
{item.path}

{item.desc}

))}

Most setups rely on promises:{' '} trust the model provider's policies, trust the application code, trust that the agent respects boundaries. Greywall replaces trust with enforcement. Constraints are applied at the kernel level, below anything the agent or model can circumvent.

) }