# Specialized agent guards block up to nine in ten safe actions when tool calls arrive as shell commands, ToolRosetta finds

Frontier and strong open-weight models used as guards gave the same verdicts whether an action appeared as JSON, Python or shell, while four of five purpose-built agent guards did not.

- Published: 2026-10-20T06:12:48.347Z
- Canonical: https://polylog.news/ai/2026-10-20/specialized-agent-guards-block-up-to-nine-in-ten-safe-action
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [arXiv cs.CR](https://arxiv.org/abs/2610.24055), [arXiv cs.CR](https://arxiv.org/abs/2610.24052), [arXiv cs.CR](https://arxiv.org/abs/2610.24058), [NVIDIA Blog](https://blogs.nvidia.com/blog/media-agentic-security-european-public-service-broadcasting-global-content-production/)

Step-level guards for tool-using agents are usually evaluated on clean JSON tool calls. Deployed coding agents, however, act through executebash and executecode tools whose payloads are plain strings. ToolRosetta, a new paper also listed fo…

This story is for subscribers. Read it in full at https://polylog.news/ai/2026-10-20/specialized-agent-guards-block-up-to-nine-in-ten-safe-action (subscription information: https://polylog.news/pricing).