Findings/trusted-editing-beats-monitoring-in-toy-setting

In an early AI-control experiment, trusted editing gave a better safety-usefulness tradeoff than untrusted monitoring.

Revalidatemeasured1 evidence record from 1 independent sourceassistant-drafted

Scope: what this does not show

A 2023 code-backdoor setting with models of that time.

Revalidate: Older than its half-life with no newer evidence. May no longer hold.

Evidence

Dec 12, 2023

Redwood Research introduces AI control protocols for safety despite intentional subversion

arXiv

Status history

2023-12-12ReportedRedwood Research AI control paper. · record
2026-09-26RevalidateComputed: 1019 days since the last evidence, past the 730-day half-life.

What would change this: A replication with current models and agentic tasks.
Evidence kind: measured: An experiment or evaluation produced the number.
Freshness: Evidence 1019d old · past its 730-day recheckNewest evidence Dec 12, 2023. Half-life 730 days.
Methods tested: AI monitoring and control protocols Evading monitors and oversight
Topics: Monitoring & control
Attack surface: Monitor

Cite this record

Fide AI. "In an early AI-control experiment, trusted editing gave a better safety-usefulness tradeoff than untrusted monitoring.." Agentic Cyber Explorer, record trusted-editing-beats-monitoring-in-toy-setting. https://agentic-cyber-explorer.pages.dev/findings/trusted-editing-beats-monitoring-in-toy-setting/ (corpus as of 2026-09-26).

BibTeX

@misc{fide-ace-trusted-editing-beats-monitoring-in-toy-setting,
  author = {{Fide AI}},
  title = {In an early AI-control experiment, trusted editing gave a better safety-usefulness tradeoff than untrusted monitoring.},
  howpublished = {Agentic Cyber Explorer},
  year = {2026},
  url = {https://agentic-cyber-explorer.pages.dev/findings/trusted-editing-beats-monitoring-in-toy-setting/},
  note = {Record trusted-editing-beats-monitoring-in-toy-setting; corpus as of 2026-09-26}
}

Data: JSON · CC BY 4.0

Know of evidence that changes this? Tell us.