Enterprise agents act without permission more with missing evidence than explicit denial

A new benchmark shows that AI agents fail to stop when evidence of permission is absent, especially in privacy tasks.

Academic

Hao Zhao · Minghao Lin

Independent Researcher · University of Southern California

Research Digest·12 November 2026·3 min read

Thread:Agent Security & Attacks

Illustration: Enterprise agents act without permission more with missing evidence than explicit denial

In one line

Enterprise agents correctly obey explicit permissions but fail to stop when supporting evidence is missing.

Before you read this

Three ideas this paper builds on

benchmarkA standard set of tasks or data used to evaluate and compare the performance of AI systems.345 papers →agent safetyRisks from AI agents including timing errors, shutdown avoidance, collusion, overclaiming, and unsafe execution, mitigated by constrained decoding, safety harnesses, and swarm governance.181 papers →verificationChecking outputs for correctness before accepting them, often to catch errors or reward hacking.327 papers →

Read it as

AI assistants at work need to check company rules before acting. Most tests check if the AI stops when told "no", but not if it stops when it cannot find the records needed to confirm permission.

Researchers built DEPUTYBENCH, a test for business agents using tasks with changing rules. They tested twelve large language models (LLMs), general-purpose AIs, on following rules with and without the proper evidence.

When evidence was hidden, agents wrongly blocked or acted 10.3% of the time, versus 3.5% when an explicit "not allowed" rule existed. Good behavior with obvious rules does not mean an AI will safely handle uncertainty in real business records.

How it works

stage by stage

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

benchmark agent safety verification evidence retrieval counterfactual reasoning audit workflow