Agent runtime guardrails

Dreadnode's best LLM scope gate missed one in ten violations across 4,897 tool calls

Dreadnode benchmarked eight judges as pre-execution gates for offensive agents. The top performer reached human range on 4,897 real tool calls, which still leaves a tenth of out-of-scope actions arriving at whatever sits on the other end.