CLAWX applies a least-privilege model across 58 catalogued marketplace skills. Each skill's memory blocks are encrypted with that skill's own AES-256-GCM key, and a read is refused when the requesting skill is not the owning skill (for example, a Smart Home skill asking for a 1Password skill's memory). The requesting skill is declared by the caller, so this depends on the agent runtime authenticating which skill is calling; it is an in-process check, not OS or hardware isolation.
Simulate an adversarial scenario: run a Smart Home process thread that attempts to read the protected memory buffer of a 1Password secret vault or GitHub developer token.
Each skill operates under strict least-privilege tool scopes, memory partitions, and outbound network boundaries.
CAIN-42 MCPGate