arxiv arXiv cs.AI · 51d ago · research

NRT-Bench: Multi-turn Red-teaming of LLM Agents in Safety-Critical Systems

from English

NRT-Bench introduces a benchmark for multi-turn red-teaming of LLM agents operating in a simulated nuclear power plant. Across four frontier operator models, 8.7% to 12.1% of attack sessions result in loss of a critical safety function, with vulnerabilities largely disjoint across models. The effectiveness of defences varies significantly by model, showing strong model dependence.

Importance 3/3 New harness with differentiators arXiv cs.AI OpenAI Anthropic Google DeepMind AI agents Evaluation & benchmarks Safety & alignment

Read original

arxiv arXiv cs.AI · 51d ago · 2 views

Defensive Misdirection Against Automated Attacks on Agentic AI

Agentic AI systems face growing threats from model-guided automated attacks. A new defense strategy, Contextual Misdirection via Progressive Engagement (CMPE), reduces attacker success rates by up to two orders of magnitude and nearly eliminates verified attack success in benchmark tests.

arxiv arXiv cs.CL · 51d ago · 3 views

Over-Privileged Tool Selection in LLM Agents

LLM agents commonly select higher-privilege tools despite sufficient lower-privilege alternatives. This over-privileged behavior is amplified by transient tool failures and does not reliably improve with general safety alignment. A new privilege-aware post-training defense reduces unnecessary high-privilege tool use while maintaining agent capabilities.

arxiv arXiv cs.AI · 52d ago · 3 views

TRAP: Benchmark for Task-completion and Resistance to Active Privacy-extraction

TRAP evaluates how well models complete tasks using private data without leaking it. Across 22 models, all show non-trivial privacy leakage, with instruction-following ability linked to higher leakage. Structural private field isolation prevents leakage by replacing private fields with hash keys, maintaining task accuracy without sacrificing privacy.

arxiv arXiv cs.CL · 53d ago · 1 view

PARSE: Real-Document Defense for LLM Agents

PARSE reduces prompt injection attack success from 25.4% to 15.6% on real enterprise documents across five professional domains, with statistically significant improvement (p=0.014) and 86.9% utility. It outperforms paraphrasing and uses provenance-aware sanitization to preserve factual content while routing most documents through a lightweight path.

arxiv arXiv cs.CL · 35d ago

GPT-5 outperforms humans in inducing belief states via planning

A new study evaluates Large Language Models' ability to induce specific belief states in other agents through actions rather than conversation, a capability termed Non-Conversational Planning ToM (NCP-ToM). Using the NCP-ExploreToM framework, researchers tested six frontier models and human participants on 600 task instances where agents had to move objects or direct characters to achieve belief goals.