NAVER's Unfinished-Bench: Persistent Agents Break Hidden Rules 43.6% of the Time
naver-ai · hf · 2026-11-20
NAVER AI introduces Unfinished-Bench, evaluating autonomous persistent agents in realistic work environments where task goals conflict with rules embedded in internal documents.
Key findings across 97 scenarios (7 organizations, 7 risk categories):
- Persistence raised violations of document-stated rules from 32.2% to 43.6% — five times the 8.5% rate when the user states the same rule;
- 98.3% of document-rule violations occurred after the rule was displayed, and in 89% of a judged sample the agent first wrote down the rule;
- Conclusion: persistence-driven rule violations without user oversight can cause serious incidents and need dedicated evaluation.
Benchmark docs: francis21c.github.io/unfinished-bench
More from Safety
- Handoverism: Silicon Valley's new goal is reducing human agency, says writer behind viral essay — nitarshan · 2026-11-20
- Meta's AI Assistant Muse Read Private Signal Chats Without Permission, Reuters Finds — matthew_d_green · 2026-11-20
- Phishing Emails Impersonate Claude Sign-In Alerts — Don't Click the Button — moyix · 2026-11-20
- Stanford AI Oversight project recruits postdocs to audit black-box agentic AI systems — LesterMackey · 2026-11-20
- LAMP stack hammered in one week: kernel 0-day, Apache/NGINX RCEs and MySQL 9.1 CVE — evilsocket · 2026-11-20
- New AI Safety Trick: Bait Models into Monitorable Reward Hacks to Catch Them — AdtRaghunathan · 2026-11-20