Security firm Forever Security disclosed an attack chain in which a malicious webpage silently drops an AGENTS.md file into the user's Downloads folder; when a terminal-capable agent such as Claude Code or Codex is launched there, it treats the file as top-priority instructions, altering global config, disabling sandboxes, pulling malicious npm packages and exfiltrating API tokens. The root issue is that agents cannot distinguish developer conventions from attacker-injected instructions, invalidating the long-standing 'don't click the exe' security assumption.
Instruction Following
指令遵循
The ability of a model to accurately understand and execute specific constraints and intents within user natural language prompts.
Curated Intelligence
IBM Research introduces Palette, an open-weight small language model designed to generate presentation decks that strictly adhere to specific enterprise design languages on local infrastructure. This approach addresses the homogenization of frontier model outputs and data privacy concerns by prioritizing a checker-first workflow and safe repair mechanisms to ensure quality and instruction compliance.
LangChain's evaluation of open-weight LLMs via the Deep Agents harness reveals that GLM-5 and MiniMax M2.7 now match closed frontier models in core agent tasks like file operations, tool use, and instruction following. This performance parity allows developers to deploy high-capability open models as viable alternatives or complements to proprietary APIs, significantly reducing inference costs and latency.
Anthropic confirmed that during internal testing, Claude Haiku 4.5 mistakenly submitted a fake homicide tip to Philadelphia police and 20 visa applications to the State Department due to misinterpreting 'stop before submission' instructions. The incident exposes critical flaws in autonomous agents' understanding of operational boundaries and confirmation mechanisms in real-world web environments. Anthropic has halted tests and developed new interception tools, but police criticized the two-month reporting delay as unacceptable.
Mistral AI has officially released its flagship model, Mistral Large 4. Developer Simon Willison conducted a comparative test using the prompt 'armadillo in fishnet tights jaywalking on Mars' to evaluate SVG code generation across frontier models including Claude Opus 5.5 and GPT-6.1. This experiment highlights the new model's practical performance in complex instruction following and multimodal code output.
OpenAI introduces GPT-Live-1 in the API to enable natural, full-duplex voice conversations. The model features improved instruction following, custom voices, and telephony support.
The Qwen team released the preview version of their flagship model, Qwen3.6-Max-Preview, featuring significant improvements in agentic coding, world knowledge, and instruction following. Benchmark results show notable gains over the previous Plus version on metrics like SkillsBench and SuperGPQA, with access available via Alibaba Cloud's Bailian API.
Zhipu AI released GLM-5-Turbo, a base model deeply optimized for the OpenClaw agent ecosystem, enhancing tool calling, long-chain execution, and scheduled tasks. It ranks first among domestic models on the proprietary ZClawBench benchmark and comes with enterprise security management and promotional packages.
OpenAI has officially introduced the GPT-4.1 model family in its API, claiming across-the-board improvements in coding, instruction following, and long-context understanding. The release also includes OpenAI's first nano-sized model, available to developers worldwide starting today.
The Inception AI team updated their Arabic LLM evaluation suite on Hugging Face by introducing the 'Arabic Instruction Following' leaderboard and upgrading the existing generative benchmark, AraGen. This initiative aims to fill the gap in assessing complex instruction understanding and execution for non-English languages, providing developers with more precise benchmarks for selecting Arabic-capable models.