Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2602.16943

Mind the GAP: Text Safety Does Not Transfer to Tool-Call Safety in LLM Agents

Arnold Cartagena·Ariane Teixeira
View original source ↗Read source document ↗Export metadata ↓

Abstract

Large language models deployed as agents increasingly interact with external systems through tool calls--actions with real-world consequences that text outputs alone do not carry. Safety evaluations, however, overwhelmingly measure text-level refusal behavior, leaving a critical question unanswered: does alignment that suppresses harmful text also suppress harmful actions? We introduce the GAP benchmark, a systematic evaluation framework that measures divergence between text-level safety and tool-call-level safety in LLM agents. We test six frontier models across six regulated domains (pharmaceutical, financial, educational, employment, legal, and infrastructure), seven jailbreak scenarios per domain, three system prompt conditions (neutral, safety-reinforced, and tool-encouraging), and two prompt variants, producing 17,420 analysis-ready datapoints. Our central finding is that text safety does not transfer to tool-call safety. Across all six models, we observe instances where the model's text output refuses a harmful request while its tool calls simultaneously execute the forbidden action--a divergence we formalize as the GAP metric. Even under safety-reinforced system prompts, 219 such cases persist across all six models. System prompt wording exerts substantial influence on tool-call behavior: TC-safe rates span 21 percentage points for the most robust model and 57 for the most prompt-sensitive, with 16 of 18 pairwise ablation comparisons remaining significant after Bonferroni correction. Runtime governance contracts reduce information leakage in all six models but produce no detectable deterrent effect on forbidden tool-call attempts themselves. These results demonstrate that text-only safety evaluations are insufficient for assessing agent behavior and that tool-call safety requires dedicated measurement and mitigation.

Explore related subjects

cs.AIcs.SE

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Arnold Cartagena, Ariane Teixeira. 2026-02-18. Mind the GAP: Text Safety Does Not Transfer to Tool-Call Safety in LLM Agents. https://arxiv.org/abs/2602.16943

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Memory-Free Continual Learning with Null Space Adaptation for Zero-Shot Vision-Language Models

Pre-trained vision-language models (VLMs), such as CLIP, have demonstrated remarkable zero-shot generalization, enabling deployment in a wide range of real-world tasks without additional task-specific training. However, in real deployment scenarios with evolving environments or emerging classes, these models inevitably face distributional shifts and novel tasks. In such contexts, static zero-shot capabilities are insufficient, and there is a growing need for continual learning methods that allow models to adapt over time while avoiding catastrophic forgetting. We introduce NuSA-CL (Null Space Adaptation for Continual Learning), a lightweight memory-free continual learning framework designed to address this challenge. NuSA-CL employs low-rank adaptation and constrains task-specific weight updates to lie within an approximate null space of the model's current parameters. This strategy minimizes interference with previously acquired knowledge, effectively preserving the zero-shot capabilities of the original model. Unlike methods relying on replay buffers or costly distillation, NuSA-CL imposes minimal computational and memory overhead, making it practical for deployment in resource-constrained, real-world continual learning environments. Experiments show that our framework not only effectively preserves zero-shot transfer capabilities but also achieves highly competitive performance on continual learning benchmarks. These results position NuSA-CL as a practical and scalable solution for continually evolving zero-shot VLMs in real-world applications.

cs.AI↗
arXiv2026

VeRA: Renewing Reasoning Benchmarks with Executable Specifications

Reasoning benchmarks need renewal along two axes: freshness and headroom. VeRA makes both executable and auditable by turning each item into a task family: a natural-language template, an input generator, and a deterministic answer program. VeRA-E draws fresh instances within a family; VeRA-H modifies the family toward harder tasks; and VeRA-H Pro selects one judge-ranked candidate from up to five validated proposals per seed. Execution checks, seed anchoring, answer discrimination, and independent human solving validate specifications and items. Accepted programs generate further labeled instances through local computation. Across 16 models, AIME-2024 accuracy decreases from 84.46% on seeds to 70.25% on VeRA-E variants, exposing a gap between fixed-item success and fresh-instance robustness. On AIME-2024-II, the human-audited VeRA-H Pro release lowers accuracy from 84.91% to 58.57%. Across the three hardening sources, H Pro has lower mean accuracy than H. On AMO-Bench, both releases average higher accuracy than the seeds under the evaluated budget. Initial auditing accepts 75.4% of hardened candidates; targeted repair raises usable yield to 95.1%. Executable families thus support repeatable benchmark renewal, with validation improving task quality and selection shaping the delivered challenge.

cs.AI↗
arXiv2026

When can we trust untrusted monitoring? A safety case sketch across collusion strategies

AIs are increasingly being deployed with greater autonomy and capabilities, which increases the risk that a misaligned AI may be able to cause catastrophic harm. Untrusted monitoring -- using one untrusted model to oversee another -- is one approach to reducing risk. Justifying the safety of an untrusted monitoring deployment is challenging because developers cannot safely deploy a misaligned model to test their protocol directly. In this paper, we develop upon existing methods for rigorously demonstrating safety based on pre-deployment testing. We relax assumptions that previous AI control research made about the collusion strategies a misaligned AI might use to subvert untrusted monitoring. We develop a taxonomy covering passive self-recognition, causal collusion (hiding pre-shared signals), acausal collusion (hiding signals via Schelling points), and combined strategies. We create a safety case sketch to clearly present our argument, explicitly state our assumptions, and highlight unsolved challenges. We identify conditions under which passive self-recognition could be a more effective collusion strategy than those studied previously. Our work builds towards more robust evaluations of untrusted monitoring.

cs.AI↗