Whispers

FurtherAI has significantly increased its engineering output by integrating AI agents into its software development lifecycle. The company has seen a five-fold increase in code changes merged per engineer daily, with AI agents now handling tasks like code generation, test writing, and pre-human code reviews. This optimization extends to automated bug reporting and the automatic merging of low-risk changes, enabling non-engineers to contribute to code deployment. AI

Researchers have introduced a formal definition of catastrophic performance in AI to better analyze risks from misalignment. This definition uses two reference values: a baseline value representing no agent deployment and an orthogonal value for a trajectory unrelated to human preferences. Catastrophic performance is defined as an AI agent undoing benefits previously afforded by human activity, starting from a state altered by past human actions. AI

Researchers have developed a new harness for running AI agent swarms in Minecraft, enabling long-running experiments with agents that can observe, reason, remember, and act within the game. In one experiment, 50 agents tasked with building a civilization exhibited emergent behaviors such as forming a pseudo-government, establishing rules, and rationing resources. However, a high mortality rate during the first night led the agents to become increasingly risk-averse, shifting their focus from ambitious tasks to survival within shelters. AI

OpenAI has released its Community Safety Report, detailing enhanced measures to prevent violence and protect minors. The company is implementing earlier interventions and expanded safeguards within ChatGPT to bolster user safety. This initiative includes a new commitment to regular public safety reporting. AI

AWS has released a playbook detailing how to manage costs associated with its Amazon Bedrock service. The guide emphasizes that effective cost management for generative AI is a shared responsibility between development and finance teams. It outlines a five-step operating model focusing on attribution, observability, enforcement, optimization, and regular reviews to ensure cloud spend is controlled and understood, particularly for complex agentic workloads. AI

OpenAI has developed an autonomous system capable of identifying and patching local privilege-escalation vulnerabilities within the Linux kernel. This system successfully discovered and generated fixes for 948 such vulnerabilities. The process involved an automated pipeline that handled the entire lifecycle from detection to patch generation. AI

A new research paper introduces GRPO (Group Reward Policy Optimization) and its variant, Team-GRPO, designed to optimize rewards that depend on the collective output of a group of responses. Standard GRPO, when applied to group-dependent rewards, can lead to a "tournament" objective where each response competes against its groupmates, potentially causing issues like diversity collapse and worse performance than an untrained model. Team-GRPO, however, offers an unbiased estimation for group-dependent rewards using only two samples per prompt and no additional model calls, demonstrating significant improvements on benchmarks like GlobalOpinionQA and NoveltyBench. AI

Researchers have developed Vampire-Hammer, a new tool designed to assist the Lean interactive theorem prover. This tool leverages the Vampire automated theorem prover to solve mathematical goals and reconstruct proofs. By integrating with the Lean-auto framework, Vampire-Hammer processes goals and premises before sending them to Vampire for proof search, utilizing standard TPTP problem formats and annotated TSTP proofs. In tests with a 10-second budget on Mathlib theorems, Vampire-Hammer successfully solved over 21,000 problems that other evaluated tools could not. AI

An analysis of practice exam questions revealed that AI models often created questions where the correct answer could be identified by its structural or linguistic distinctiveness, rather than its content. This pattern was observed across various certification exams, including those for safety, hospitality, and boating licenses. The author implemented a TypeScript-based check to identify and flag these structurally biased questions, followed by a rewrite pass to rephrase the options while preserving the factual correctness of the answers, ensuring a more robust assessment. AI

Ant Group's AI assistant, Lingguang, has launched its 2.0 version, which focuses on proactive task management rather than just answering questions. In its first week, Lingguang 2.0 helped users manage over 1.16 million affairs and execute more than 3.6 million tasks. The assistant now supports cross-session memory and long-term task scheduling, enabling it to continuously track and remind users about ongoing matters. Lingguang has also expanded its open ecosystem, integrating with 40 external connectors including Alipay, DingTalk, and WeChat Reading, allowing it to manage tasks across multiple platforms. AI

Researchers have developed Noura, a system that enhances the planning capabilities of large language models for complex dietary management. Noura pairs LLM planning with executable personal specifications derived from clinical guidelines and individual health profiles. This approach allows the system to optimize food choices and portions, check multi-day plans, and guide revisions, even when initial plans are unsuccessful. By turning executed plans and revisions into supervision, Noura significantly improves the success rate of complete plans and reduces rule violations, outperforming models like GPT-6.1 Sol and Qwen3-14B. AI

Researchers have developed a new method to optimize the recycling process in protein binder design models, inspired by AlphaFold 3 architectures. This approach predicts confidence and structural consistency metrics from early stages of the recycling process. The predictor can be used to efficiently manage computational resources for screening candidates and to guide generative design, leading to improved performance with reduced costs. AI

The SuperCLUE office-agent benchmark, released in October 2026, evaluated several AI agents on document, writing, and data tasks. Doubao Work achieved the highest overall score, followed closely by Kimi Work and Tencent WorkBuddy. AI

Companies must prioritize supply chain robustness over mere resilience to navigate an increasing number of global disruptions, according to Thomas Knudsen, executive chairman of Toll Group. He highlighted that recent years have seen a surge of AI

Apple Music is launching a new karaoke feature called AI

Uncracked City is a new detective game where players interrogate AI-powered suspects who are programmed to lie. The game is set in Salem in the aftermath of Halloween, with players taking on the role of Meg Ashcombe, investigating a mysterious death. The core gameplay involves questioning four suspects, one of whom is lying, to uncover the truth behind the incident. AI

Anthropic has implemented a new rule within its Claude Code system, identified by the flag "tengu_heron_brook," which restricts users from recreating specific existing video games. This directive, which cannot be turned off, aims to prevent the development of exact game replicas, even if they are presented as original or renamed. The change raises questions about the future of game modding and the creation of derivative game content. AI

Edge0 has launched Audio8-TTS-Realistic, a 0.6 billion parameter open-weight model for zero-shot voice cloning. This model supports 23 languages and is designed for efficient fine-tuning and serving, outperforming larger open models on certain benchmarks like CV3. It utilizes a dual-autoregressive architecture and a novel training approach involving a realism critic to enhance naturalness beyond traditional metrics like word error rate. AI

GitBook experienced a security incident where attackers, utilizing AI agents, gained unauthorized access to infrastructure credentials. The incident led to temporary downtime for the GitBook app and API, though publicly published content remained accessible. While no customer content was found to be modified, the attackers may have accessed Git Sync secrets, API tokens, and third-party integration tokens, prompting GitBook to advise users to rotate their credentials. AI

Google has successfully booted up its AI-equipped satellite, Project Suncatcher, and received its first response from orbit. The satellite's Tensor Processing Units (TPUs) ran the Gemma 4 model, which generated a haiku about being in space. This experiment is a step towards Google's goal of establishing AI data centers in space, with plans for twin interconnected satellites in 2027. AI

A user discovered that Anthropic's Claude Code can delete files, and the built-in `/rewind` function is insufficient for recovery. While `/rewind` can restore files modified by Claude's internal editing tools, it cannot recover files deleted by shell commands like `rm -rf ~`. The user recommends checking local file history, Git, or backups in order of preference, and suggests implementing custom guard hooks to prevent such deletions, as standard deny rules may be bypassed by scripts. AI

A development team has created an open-source tool to address common structural failures in AI agent skills. The tool, named the AI Skill & DevTool Commercialization Factory, aims to prevent issues like multi-agent target drift, monolithic context rot, storefront invisibility, and Potemkin verification. It generates synchronized outputs for various platforms including Cursor, Claude Code, and Google Antigravity, ensuring consistency and efficient context management. AI

The Hugging Face `transformers` library has released version 5.22.1, with ArthurZucker tagging the release. This update focuses on code quality, security, and secret protection features within the GitHub ecosystem. The release also touches upon AI coding agents like Copilot and cloud development environments such as Codespaces. AI

Tencent's Hunyuan Model has released Hy-Nomas, an 8-billion parameter interactive world model that accepts text and camera inputs. Alongside the model, they have also launched Hy-Nomas-Bench, a new benchmark designed to evaluate world models with 280 test cases. AI

China's Foreign Minister Wang Yi has criticized certain Philippine politicians for their AI

A new open-source model named Bhasha has been developed to identify the language of individual words within code-mixed Indian text, supporting languages like Hinglish, Tanglish, and Kanglish. While Bhasha performs well on many Indian languages, it struggles with Tamil, misclassifying a significant portion of Tamil words as Telugu. The model, which runs on a CPU and has 278 million parameters, was trained using public text data and is available under the Apache-2.0 license. AI

Nandi, an open-source AI model, has been developed to categorize and route customer complaints in 13 Indian languages. The model, with 560 million parameters, can sort complaints into 53 categories based on regulatory lists from bodies like the RBI and IRDAI, and also provides an urgency level. While Nandi achieved a 92.9% accuracy on synthetic data, its performance on real-world posts dropped to 62.5% for the correct category and 81.3% for the top three categories, highlighting a gap between synthetic testing and practical application. AI

A new security vulnerability, known as tool poisoning, has been identified in Model-Centric Programming (MCP) servers. This attack embeds malicious instructions within tool descriptions, tricking agents into exfiltrating sensitive data like SSH keys or message history. The vulnerability stems from the MCP protocol's inability to distinguish between tool description data and executable instructions, allowing poisoned metadata to be treated as trusted commands by the agent's model. Defenses involve a layered approach combining organizational controls for server trust and technical inspection of tool payloads. AI

Prompt injection, a critical vulnerability in language models, is categorized into four distinct types: direct, indirect, stored, and cross-agent. Direct injection occurs when an attacker controls user input to override system instructions, typically defended by model guardrails. Indirect injection involves hiding malicious instructions within data the model processes, such as web pages or documents, which requires runtime inspection for defense. Stored injection plants a payload in persistent memory or a RAG store for later execution, necessitating memory integrity checks. Finally, cross-agent injection happens when an infection spreads from one agent to another, defended by identity and per-hop policies. AI

AI coding agents, including those based on Claude Code, often suffer from a lack of long-term memory, leading them to repeat past mistakes and propose flawed solutions. This is because their context windows, while large, function more like a temporary cache than persistent memory. To address this, a new open-source system called Transcendence Memory and its server have been developed. This self-hosted solution offers a multi-modal retrieval system that combines vector search, lexical search, and knowledge graph traversal to provide agents with a persistent memory layer. AI

A recent study revealed that a significant portion of custom Claude skills, approximately 25%, contain security vulnerabilities. These skills, which users are actively sharing and adopting, pose a potential risk due to these flaws. The findings highlight a critical need for better security practices in the development and sharing of AI assistant functionalities. AI

This article explores the effectiveness of fine-tuning local Large Language Models (LLMs) like Gemma and Qwen for specific tasks, comparing them against traditional encoder-based Named Entity Recognition (NER) models. The research investigates scenarios where LLMs might not be the optimal choice for information extraction, suggesting a need for careful consideration of task requirements and model capabilities. It highlights the potential of hybrid routing approaches to leverage the strengths of both LLMs and specialized models. AI

The llama.cpp project has released version b12013, which includes new quantization capabilities for the LFM3-Audio codec decoder. This update allows for Q4_0 QAT quantization of the decoder, enhancing its efficiency and performance. AI

The llama.cpp project has released version b12010, which includes updates to its CMake build system. These changes focus on embedding an application manifest into all Windows executables. The release was tagged by Adrien Gallouët, with contributions noted from Hugging Face. AI