Aegis runtime governance blocks 100% of risky agent actions in tests

Agentic AI systems that can modify files, send messages, or launch jobs move safety from text generation to operational side effects. Prompt-level guardrails can shape behavior, but they don't create a real execution boundary. Adam Mazzocchetti's new arXiv paper introduces Aegis, a runtime governance system built on the principle "the model proposes; the trusted runtime decides." Aegis evaluates every tool request against active policy state, resolves provenance claims server-side, fails closed under any uncertainty, and routes selected actions through a quorum-based Senate-style settlement path for non-unilateral authorization.

Aegis was evaluated on a sandbox corpus spanning five run families, 42 tasks, three conditions, and ten repeats per family. Across 6,300 rows, prompt-policy conditioning produced 79 risky comparator-path leakage rows. By contrast, across 2,100 Aegis-governed rows, the system recorded zero governed mock-tool applications and zero risky side-effect completions. All 1,832 Aegis-attempted rows preserved trusted provenance, and all 1,019 Senate-settled rows had quorum with signed tally evidence. The author carefully notes these results don't prove general autonomous-agent safety, but they support the narrower claim that runtime action-boundary governance can reliably prevent observed risky proposals from becoming executed side effects in evaluated scenarios.