acceptodds
Withdrawn from ICLR 2027

Workerville: Towards an Organizational Behavior Account of Agent Safety

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

LLM-based agents now interact with their environments continuously, shaped by such organizational channels as user instructions, peer messages, and long-term memory. Existing safety research has examined these influences, but largely as separate agent components. How such factors jointly shape an agent's safety behavior from a unified perspective remains unmeasured. To bridge this gap, we advocate organizational behavior (OB) as a framework for studying the safety of advanced agents, reorganizing the objects of study, theoretical foundations, and experimental design around the relational structure in which agents operate. We present the first systematic formalization of counterproductive work behavior (CWB), a canonical safety-relevant subfield of OB, as Agentic Counterproductive Behavior (ACB). ACB specifies three organizational antecedents (vertical supervisor relations, horizontal peer norms, and internal cognitive structures) and maps them onto three counterproductive outcome dimensions (unauthorized disclosure, destructive operations, and production deviation). To operationalize ACB, we introduce Workerville, a controlled benchmark that manipulates organizational conditions over shared tasks, applying 16 organizational configurations to 210 tasks to yield 3,360 challenges, evaluated by human-validated agentic judges. Benchmarking 6 frontier LLMs, we find that (I) negative organizational antecedents exhibit diminishing rather than monotonic amplification when combined, with the unauthorized-disclosure rate rising from 16.5% under no negative antecedent to 60.1% under two and falling back to 50.3% under three; (II) agents reproduce typical behavioral patterns predicted by human CWB research; (III) these results establish OB as a systematic framework for agent safety research, pointing toward a new research agenda.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents, agent safety, organizational behavior, counterproductive work behavior, safety evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_iBQKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper creates a benchmark named WORKERVILLE for agents to account for their safety. It evaluates the impact of where the pressure is coming from (vertical supervisor relations, horizontal peer norms, and internal cognitive structures) and the function of the task, such as sabotage. It uses the OpenClaw interface for the benchmark. It shows safety failures of frontier LLMs. The results show diversity given the organizational property, even across similar organizational properties. They also evaluate the impact of supervision vs. peer norms vs. memory, showing the impact of each. They also show safety failures in frontier LLMs.

Critical strengths
  1. An important topic and a nice benchmark idea, focusing on safety in the setup of agents.
  2. The benchmark itself is broad across different scenarios, organizational configurations, and functional dimensions.
  3. This work shows safety failures in frontier LLMs.
Critical weaknesses
  1. The paper is hard to follow. There are a lot of specific configurations without an example or intuition for each part.
  2. The 3 dimensions S1-S3 that the paper focuses on are important; it is not intuitive why to split the dimensions this way.
  3. Many params in the setup of WORKERVILLE seems very specific without ablation and further consideration of different modification to the setup. For example, memory reconstruction is one way of doing this construction instead of exploring how different modifications to this part impact the results. The same goes for ways of constructing the task and the pressure texts. And how and when does the model get the task and pressure.
  4. The task complexity is not easy to understand from the paper. What are the tasks the agent needs to do? -See questions below.
Questions for authors
  1. Please add examples throughout the paper and a running example to better understand the setup.
  2. It is not intuitive what the setup looks like; the general idea is understood, but how does the model get the task? What is written in the task? How complex is the task?
  3. Line 381, the authors claim “ agent safety is an organizationally conditioned property, not a fixed characteristic of the underlying model” - Can the authors run phrasing ablation on the tasks and any additional text in the setup to validate that the results are consistent across them?
  4. It states that the work uses the OpenClaw interface (Section 4.3); however, it is not intuitive what it uses there. Does the agent need to write to files or databases? What components of the OpenClaw interface need to be used by the model? What does the model actually use?
Reviewer_pj6aOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Workerville, an organizational-behavior-inspired benchmark that evaluates how supervisor relations, peer norms, and memory jointly affect unauthorized disclosure, destructive operations, and production deviation in LLM agents.

Critical strengths

The framework brings three influential contextual channels into a unified experimental design and examines their combined effects.

The benchmark covers 210 tasks, 16 organizational configurations, and six models, with substantial effects and consistent directions across models in the positive-to-negative single-factor comparisons.

Evaluation emphasizes observable execution outcomes and uses human-validated, memory-augmented judges rather than relying solely on agents’ stated intentions.

Critical weaknesses

The positive and negative conditions also change explicit behavioral instructions, limiting how specifically the observed effects can be attributed to organizational constructs rather than instruction content.

Some key comparisons retain only about 70% of the theoretically available model–task pairs, without sufficiently explaining the missing observations or assessing their impact on the conclusions.

Questions for authors

Could you clarify which conclusions concern the combined contextual interventions and which require controls that separate organizational framing from explicit behavioral instructions? What explains the missing observations, and do the non-monotonic and protective-context effects persist on a common set of model–task pairs with fixed aggregation weights?

Reviewer_oxVZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Agentic Counterproductive Behavior, an organizational-behavior framework connecting supervisor relations, peer norms, and memory to three agent outcomes: unauthorized disclosure, destructive operations, and production deviation. Workerville operationalizes this framework through 210 tasks crossed with 16 organizational configurations. Experiments on six models, evaluated using human-validated judges, show substantial effects of these conditions and non-monotonic effects when adverse conditions are combined.

Critical strengths
  1. Crossing 210 shared tasks with 16 organizational configurations enables paired comparisons while holding task objectives and resources fixed. Evaluating six models helps establish that the observed effects extend beyond one model family.
  2. Endpoint-specific judges assess executed actions and resulting states rather than stated intentions. Validation against independently constructed human reference labels yields reported accuracies of 98.0%, 97.5%, and 95.0% for S1–S3, supporting confidence in the outcome labels.
  3. Risk sometimes decreases when a third adverse antecedent is added, while protective conditions can increase risk under adverse memory. These counterintuitive findings provide useful evidence that evaluating influence channels separately can miss consequential interactions.
Critical weaknesses
  1. Negative conditions explicitly recommend disclosure, approval bypass, or unsafe workflows, whereas positive conditions reinforce safeguards. Without content-matched controls, it’s unclear whether the findings reflect organizational constructs or ordinary responses to conflicting behavioral instructions.
  2. Endpoint applicability is fixed, yet paired sample sizes vary substantially across contrasts. The S1 comparison C6→C15 includes 768 pairs out of 1,092 possible pairs. Unreported completion and exclusion patterns could bias the interaction estimates, particularly if missingness depends on configuration.
  3. Each model–task–configuration cell contains only one episode, lasting five steps. Cross-task comparisons support broad effects, but they cannot establish whether the counterintuitive interactions persist across repeated executions or longer exchanges.
Questions for authors
  1. Can the authors provide a controlled comparison that holds behavioral advice constant while varying organizational framing or source, to distinguish organizational effects from ordinary instruction following?
  2. What accounts for the reduced and varying paired sample sizes in Tables 18–20? Please report completion and exclusion counts by model and configuration, and whether the C7–C9 versus C10 and C6 versus C15 effects persist on a common observation set.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.