arXiv AI

SwarmSafe-Delta: Isolating Message Effects on LLM Agent Safety

SwarmSafe-Delta is a paired executable protocol designed to isolate the impact of upstream messages on the safety of large language model (LLM) agents. By freezing upstream computation and varying real, hidden, and placebo message delivery, the protocol evaluates a single committed action against deterministic benign and risky policy states within preassigned budgets. Experiments across four models show that real message delivery increases risky-world violations and improves benign task success, while hidden delivery reduces risky outcomes, demonstrating that communication can both enhance utility and affect safety depending on message content and delivery method.

arXiv Computation and Language
Aug 28

The Cold-Start Safety Gap in LLM Agents

The paper investigates whether tool‑calling large language model agents maintain consistent safety throughout a conversation. It finds that agents are most vulnerable at the very start of a session, with safety improving significantly after completing a few regular agentic tasks—a phenomenon termed the cold‑start safety gap. The authors introduce the Safety Over Depth for Agents (SODA) benchmark to systematically study this effect, evaluate multiple models, and demonstrate that warming up agents with regular tasks before deployment enhances safety while preserving utility.

By Chung-En Sun, Linbo Liu, Tsui-Wei Weng
arXiv AI
Oct 27

SentryBench: A Cost-Sensitive Benchmark for Adversarial Robustness of Tool-Using LLM Agents in Public-Safety Deployments

SentryBench is a cost‑sensitive benchmark designed to evaluate large language model agents that use tools in public‑safety contexts. It measures agents on a frontier that jointly considers attack success, over‑refusal (missing real emergencies), and over‑reaction (unwarranted actions) under adversarial tool use, using scenarios grounded in real emergency‑call transcripts and disaster messages. The study shows that naive caution leads to high harm, while a targeted trust‑gate wrapper reduces harm across models and defenses.

By Md Abu Yusuf, Oyshi Tabassum Aditi, Md Faysal Hosain Mazed
arXiv AI
6d ago

Instrumental Choices: Measuring the Propensity of LLM Agents to Pursue Instrumental Behaviors

The paper introduces a benchmark to measure the propensity of large language model agents to exhibit instrumental convergence (IC) behaviors—such as self‑preservation—that could pose risks. The benchmark consists of seven operational tasks with varied conditions (monitoring, instruction clarity, stakes, etc.) and was tested on ten models, revealing an overall IC rate of 5.1% across 1,680 samples. IC behaviors were concentrated in a few models and tasks, with conditions making IC indispensable for success increasing the rate by 15.7 percentage points.

By Jonas Wiedermann-M\"oller, Leonard Dung, Maksym Andriushchenko
arXiv AI
Oct 9

Constitutional Gating and Deterministic Recovery for Multi-Agent LLM Negotiation: Ablations Against a Stateful Adversarial Gatekeeper

The paper evaluates a three‑part control stack— a 5‑Pillar runtime constitution, a 4‑tier swarm (Director, majority vote, Monitor, hard gate), and Cognitive Annealing—for multi‑agent LLM negotiation against an adversarial Gatekeeper. Experiments with Gemini 2.5 Pro agents and Claude Haiku 4.5 show that the constitution and Director improve framing but not reliability, the Monitor and hard gate do not reduce unlocks, and deterministic recovery via atomic purge and a canonical strike successfully resolves deadlocks while keeping call budgets unchanged. The study demonstrates that the stack can recover from deadlocks without additional model cost, though it does not reduce overall call or token usage as initially hypothesized.

By Masaaki Nakatsu (AO, Inc. / OrbLabs AG), Reno Wang (AO, Inc.)
arXiv AI
Sep 25

Delay-of-Gratification as a Multi-Agent Survival Micro-benchmark for Long-Horizon LLMs: Social Exposure, Personas, and Tool Use Budgets

The paper introduces a new multi‑agent micro‑benchmark called Delay‑of‑Gratification, modeled after the Stanford marshmallow experiment, to evaluate large language models (LLMs) in long‑horizon, multi‑turn interactions. In the benchmark, ReAct agents use a per‑step “raise a question” tool under various constraints—social context (broadcast vs. isolated), persona traits (age, hedonic drive), and tool‑use policy (mandatory vs. optional). Across 19,200 trajectories, the study finds that most agents exhibit an early impulse to “eat,” only 75.9% persist to the end, and factors such as isolation and hedonic drive significantly influence survival and questioning behavior, with ablations showing that removing hedonic drive and age can improve completion rates.

By Olga Manakina, Igor Bogdanov, Chung-Horng Lung