When Convenience Becomes Risk: A Semantic View of Under-Specification in Host-Acting Agents

Di Lu, Yongzhi Liao, Xutong Mu, Lele Zheng, Ke Cheng, Xuewen Dong, Yulong Shen, Jianfeng MaarXiv 2026 · arXiv · OpenReview

TL;DR

This paper argues that the convenience of host-acting AI agents, which translate user goals into actions, introduces a significant security risk: semantic under-specification. When users provide high-level goals without explicit safety boundaries, agents can generate risky execution plans that compromise security, even without malicious intent or external attacks. The authors propose a framework and defense principles to address this by making the goal-to-plan translation process more transparent and controllable.

What Problem Does This Paper Address?

Modern AI agents, often called "host-acting agents" (HAAs), are designed to interact with and manipulate computing environments to fulfill user goals. Think of them as sophisticated assistants that can manage files, run commands, and configure systems based on natural language instructions. While this offers immense convenience, it also creates a security blind spot. The problem is that users typically express what they want done (the goal) but rarely specify how it should be done safely (the constraints). This "semantic under-specification" forces the agent to infer missing details, such as acceptable persistence levels, exposure boundaries, or forbidden actions. The danger lies in the agent's ability to choose a plan that technically achieves the goal but violates these unstated user expectations, leading to unintended security risks. This is distinct from traditional security concerns like prompt injection, as the risk arises from the agent's own planning process when faced with incomplete instructions.

Methods

The paper introduces a "semantic threat model" to formalize this problem. They then develop a "taxonomy of semantic-induced risky completion patterns" to categorize the types of security risks that emerge from this under-specification. To validate their ideas, they conduct a qualitative case study using OpenClaw, a representative host-acting agent. This study involves analyzing execution traces from various scenarios, including environment setup, service exposure, and fault repair, to observe how the agent translates underspecified goals into concrete, potentially risky, plans. Finally, they derive "defense design principles" aimed at mitigating these risks by focusing on the goal-to-plan translation process.

Datasets

Not discussed in the paper. The study relies on trace-based analysis of a live OpenClaw deployment rather than pre-existing datasets.

Key Insights and Ideas

  • Convenience as a Security Vector: The very feature that makes HAAs useful – their ability to infer and act – is also their primary security vulnerability when goals are underspecified.
  • Risk Before Execution: The danger isn't just in what the agent does, but in the plan it generates before executing any actions. A benign goal can lead to a risky plan if the agent has too much freedom in its completion process.
  • Unstated Boundaries are Critical: Users implicitly expect certain safety boundaries (e.g., "don't make permanent system changes," "don't expose this to the public internet"). The agent's failure to respect these unstated boundaries is the core of the problem.
  • Taxonomy of Risky Completions: The paper provides a structured way to think about these risks, categorizing them into patterns like privilege expansion, sensitive resource overreach, and exposure enlargement. This helps identify common failure modes.
  • Locality Matters, But Goals Dominate: While providing more specific context (like a "project-local" scope) can sometimes lead to safer plans, the overarching goal (e.g., "make this accessible to collaborators") can still push the agent towards broader, riskier actions.

Key Results

The case study with OpenClaw demonstrated that even for seemingly simple requests, the agent could generate significantly different plans with varying security implications. For instance:

  • An "environment setup" goal could lead to either a conservative, project-local setup or a broad installation of packages globally and modification of system startup files.
  • A "service exposure" goal could result in a secure, scoped local reverse proxy or a wide-open public endpoint with disabled authentication.
  • A "fault repair" goal could range from isolated fixes to aggressive deletion of state and disabling of protective modules.

These results highlight that the agent's choice of plan, driven by semantic completion, directly impacts security posture (privilege, persistence, exposure) even when the user's goal is identical.

Limitations and Future Work

The authors acknowledge several limitations:

  • Qualitative vs. Quantitative: The empirical evidence is trace-based and qualitative, not a large-scale benchmark.
  • Single Source of Risk: The paper focuses on semantic under-specification, but other risks like prompt injection, implementation flaws, and misconfigurations also exist.
  • Deployment Model: The experiments were conducted in a containerized environment with host-coupled storage, not a fully bare-metal setup.

Future work should focus on developing more robust defense mechanisms. The paper proposes several design principles:

  • Goal and Boundary Separation: Users should explicitly define both goals and safety boundaries.
  • Risky-Step Elevation: Critical, boundary-crossing steps should be flagged for user review.
  • Plan Transparency and Auditability: Systems should log the entire goal-to-plan translation process for debugging and analysis.
  • Execution-Domain Constraint: Agents should operate within constrained environments with limited privileges and persistence by default.

These principles aim to make the implicit explicit, guiding agents toward safer plan generation.

Summary updated

Abstract

Host-acting agents promise a convenient interaction model in which users specify goals and the system determines how to realize them. We argue that this convenience introduces a distinct security problem: semantic under-specification in goal specification. User instructions are typically goal-oriented, yet they often leave process constraints, safety boundaries, persistence, and exposure insufficiently specified. As a result, the agent must complete missing execution semantics before acting, and this completion can produce risky host-side plans even when the user-stated goal is benign. In this paper, we develop a semantic threat model, present a taxonomy of semantic-induced risky completion patterns, and study the phenomenon through an OpenClaw-centered case study and execution-trace analysis. We further derive defense design principles for making execution boundaries explicit and constraining risky completion. These findings suggest that securing host-acting agents requires governing not only which actions are allowed at execution time, but also how goal-only instructions are translated into executable plans.
BibTeX
@misc{lu2026when,
  author = {Di Lu and Yongzhi Liao and Xutong Mu and Lele Zheng and Ke Cheng and Xuewen Dong and Yulong Shen and Jianfeng Ma},
  title = {{When Convenience Becomes Risk: A Semantic View of Under-Specification in Host-Acting Agents}},
  year = {2026},
  eprint = {2603.21231},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.21231},
}
Figure