Brief

A new benchmark called SaaS-Bench has revealed that current AI agents struggle significantly with real-world, long-horizon tasks, with top models like Claude Opus 4.7 achieving less than 4% success rate on fully completing tasks. The benchmark uses actual SaaS systems and data, exposing four key failure modes: inability to maintain performance over extended tasks, cascading errors from single mistakes, a lack of self-checking mechanisms, and inconsistent performance across multiple runs. These findings suggest that the current paradigm for AI agents is insufficient for true automation and that software interfaces may need to be redesigned for AI agents rather than expecting them to operate human-centric interfaces. AI

A researcher has demonstrated that a smaller, open-source Turkish language model can outperform frontier models like Claude Opus 4.7, GPT-5.5, and Gemini 3.1 Pro on a specific e-commerce attribute extraction task. By fine-tuning the Trendyol-LLM-Asure-12B model with Reinforcement Learning from Human Feedback (RLHF) and using scraped product data for training, the researcher achieved statistically significant improvements in macro F1 scores. This approach offers a more cost-effective and accurate solution for specialized tasks compared to relying on general-purpose large language models. AI

Tencent Hunyuan has released its new Hy-MT2 translation model, available in three sizes (1.8B, 7B, and 30B-A3B) and supporting 33 languages. The model demonstrates strong performance, with the 7B and 30B versions outperforming many open-source models and even competing with commercial APIs like Microsoft's. Notably, Hy-MT2 shows improved instruction-following capabilities, allowing for more customized translation styles and formats, and its lightweight 1.8B version is optimized for on-device deployment with minimal storage requirements. AI

Researchers have developed a new benchmark called PPaint for image aesthetic assessment, which uses both pairwise preferences and pointwise ratings from experts. This dual-protocol approach revealed that preferences provide more consistent rankings, while ratings anchor the absolute score scale. By fusing these signals, they created a unified expert ground truth and extended the principle to training vision-language models (VLMs) without labels. A self-distillation method using this approach significantly improved an open-source VLM's aesthetic scoring capabilities, matching a closed-source model's performance with lower inference costs. AI

Researchers have developed LivePI, a new benchmark designed to more realistically assess the risks of indirect prompt injection in AI agents. This benchmark simulates real-world scenarios across various input channels like email, web pages, and chat, evaluating twelve attack families and five malicious goals. Initial tests on leading models such as GPT-5.3-Codex and Claude Opus 4.6 revealed significant vulnerabilities, with group-chat injections proving universally successful and repository link attacks causing high-severity failures. A proposed two-layer defense, combining prompt filtering and tool-call authorization, demonstrated effectiveness in blocking malicious actions without compromising agent utility. AI