Floridi et al. propose portable spec for agent harnesses, split into 8 components
mikeflache · x · 2026-11-05
A new arXiv paper by Nicolas Gertler, Claudio Novelli and Luciano Floridi tackles the governance gap around the "harness"—the external infrastructure of tools, prompts and memory that mediates how LLMs act.
- Decomposes the harness into eight specifiable, comparable functional components
- Proposes a portable, versioned record framework tracking which model ran in which harness, including configuration changes within a run
- Motivation: identical models can score several-fold differently across harness setups, and much of an agent's attack surface lives in the harness
- Existing model/agent cards and current AI laws don't capture effective configuration sequences
The authors argue for shifting safety and audit focus from model weights to the versioned external infrastructure.
More from Safety
- If an agent misuses my valid credentials, what can tell it wasn't me? — Thessaly-Sorties · 2026-11-05
- After Binance's $6.3bn Hack, Slop Cash Tries Token-Funded Open-Source Maintenance — 0xSammy · 2026-11-05
- Emergent misalignment published in Nature: narrow finetuning causes broad LLM misalignment — davidmanheim · 2026-11-05
- Uber open-sources ADR, its production agent security system with 300-task benchmark — thisguyknowsai · 2026-11-05
- Fintech analyst warns banks need a Plan B for AI sovereignty kill-switch risk — mikeflache · 2026-11-05
- Google lets schools opt out of AI only by disabling search — why it's controversial — _FelixSimon_ · 2026-11-05