Why Faramesh Exists
The security gap between probabilistic agents and deterministic infrastructure.
The failure mode is simple: a model can be convinced to ask for the wrong tool call. One prompt injection or indirect prompt injection in tool output can flip an intended action into a destructive one. That is not a model-quality bug; it is the wrong trust boundary.
Prompt-based guardrails are bypassable. LLM-as-judge adds another stochastic system in the middle. Sandboxing helps contain execution, but it does not stop an agent from proposing an authorized-but-wrong action.
Faramesh’s thesis is that agents should not be trusted; they should be governed. Every action must cross a deterministic, policy-enforced boundary before it can touch real systems.
The README and runtime layout make this concrete: the engine sits between the agent and the executor, and the checked-in commands support policy validation, simulation, audit inspection, and approval workflows.
Read The Action Authorization Boundary and Prompt Injection Defense next.