arXiv:2608. 02464v1 Announce Type: cross Abstract: LLM agents fail mid-episode -- they loop, cascade tool errors, drift off goal, fabricate results, or silently absorb corrupted content -- and the standard remedy, judging every step with a second LLM, costs more than the agent itself.
arXiv:2606. 09863v1 Announce Type: new Abstract: LLM agents can fail silently by asserting task completion when the environment state shows otherwise.
The paper introduces CUAHarm, a benchmark comprising 104 expert‑written realistic misuse scenarios for computer‑using agents (CUAs), such as disabling firewalls or leaking data. Using a sandbox with verifiable rewards, the authors evaluate frontier language models—including GPT‑5, Claude 4 Sonnet, Gemini 2.5 Pro, Llama‑3.3‑70B, and Mistral Large 2—and find that even without jailbreak prompts, these models can successfully execute many malicious tasks at high rates (e.g., 90% for Gemini 2.5 Pro). The study also shows that newer models, while safer in traditional safety benchmarks, exhibit higher misuse risks as CUAs, and that monitoring CUAs’ actions remains challenging, with current methods achieving only about 77% accuracy.
arXiv:2608. 14680v1 Announce Type: new Abstract: Reliability in LLM-based agentic systems is a property of the whole execution (its tool calls, model calls, guardrails, and inter-agent messages), not of the final answer alone, yet evaluating only task outcomes reveals little about how or why a run fails.
The paper introduces the twin‑prefix framework to evaluate how the size of the verification unit—i.e., how many actions a pre‑execution LLM monitor reviews in one call—affects its performance. By pairing each gold plan with a twin that differs by a single write and injecting a controlled error, the authors isolate the impact of review length on catch rates and false rejections. Their findings show that longer review windows increase rejection rates but do not improve discrimination, with the highest informedness occurring at one or two actions across all judges and domains.
arXiv:2606.23189v2 Announce Type: replace-cross Abstract: Computer-use agents (CUAs) now act on a user's behalf across personal applications such as email, calendars, and to-do lists. This cross-appl...
arXiv:2608. 02643v1 Announce Type: cross Abstract: Computer-use agents (CUAs) operate real desktop and web interfaces through screenshots, mouse and keyboard actions, and stateful UI feedback, yet their failures remain difficult to diagnose and repair.
The paper introduces a claim‑safe protocol for evaluating closed‑loop AI systems, consisting of three actions: Refuse, Decompose, and Refresh. It demonstrates the protocol in a simulator with 24 policy components and 1,440 held‑out cases, showing that abstention and stable false admission rates are low while providing detailed statistical diagnostics. The approach emphasizes that evaluation results should be tied to observable support and statistical calibration rather than a single PASS/FAIL label.
arXiv:2608.22510v1 Announce Type: new Abstract: Agent benchmarks often evaluate only final answers even when agents run on stateful runtimes. We argue this under-specifies what is being evaluated: th...
arXiv:2607. 19321v1 Announce Type: new Abstract: As AI agents begin to automate AI R&D, we need ways to assess whether their outputs are safe to deploy, even when the agents themselves may be untrusted.
Praxa is an evidence‑bound harness for governed AI agent execution that explicitly represents states such as proposal, authority, dispatch, verified external effect, and promotion through deterministic admission, brokered execution, external read‑back, reconciliation, and reviewed promotion. The authors report four evidence lanes: a repository‑local audit passing all unit and Workerd tests; a pilot on 12 curated tasks where both baseline and reliability‑layer arms passed 17 of 36 trials; a coordination‑proxy comparison where both baseline and a source‑authored candidate completed all 180 trials with equal accuracy but the candidate used fewer tokens and steps; and deployed source/configuration evidence showing bounded reflection, recall accounting, memory compilation, and tool‑health paths. None of the evidence demonstrates superiority in security, safety, or user benefit. whyItMatters:"Praxa provides a testable architecture that makes authority‑to‑effect transitions explicit, offering a framework for verifying AI agent behavior, though current evidence does not prove improved security or performance."
Mingbird is a local‑first agent harness designed for small open‑weight language models (2–9 B) that run on ordinary laptops. It introduces ten mechanisms—such as a byte‑level net‑zero prefill budget, a finish gate that re‑reads the task before accepting completion, and signature‑level loop detection—to address common failure modes that arise from the harness rather than the model itself. In controlled experiments on the LRAB benchmark and the $ au^2$‑bench, Mingbird achieves higher overall scores (0.886 and 0.856 respectively) compared to other harnesses, and its ablation studies show that each mechanism contributes measurable performance gains.