2026-08-28 20:51
🚨 JetBrains has alerted users of its Cadence cloud service to rotate credentials after a critical vulnerability in TeamCity was exploited. Despite disclosing the issue on July 27, an unpatched JetBrains server was compromised between August 8 and 24. The breach potentially exposed sensitive data, including credentials and source code. Users are advised to treat all affected credentials as compromised. #JetBrains #CyberSecurity #DataBreach #CloudComputing #Vulnerability
Source: The New Stack
Amanda Caswell
2026-08-28 18:31
LM Studio's recent development highlights the challenges in AI command evaluation. 🤖 Their tool, Auto Review, effectively clears 82% of commands for safety using a structure-based analysis, rather than just string checks. It builds abstract syntax trees to track command capabilities and potential risks. With over 11,651 test cases, LM Studio addresses the unique behaviors of command-line tools, ensuring more secure AI operations. 🔍 For commands that pose risks, a separate AI agent, the Shell...
Source: The New Stack
Amanda Caswell
2026-08-28 18:27
🚀 Alibaba has introduced Qwen3.8-Flash, a multimodal Mixture-of-Experts (MoE) AI model with 125 billion parameters. This model serves as a precursor to Qwen4, emphasizing performance and cost-effectiveness in coding and office tasks. By sharing its architecture early, Alibaba invites developers to explore and prepare for future developments. Qwen3.8-Flash also showcases improvements in attention, optimization, and model capacity. #AI #MachineLearning #Alibaba #TechNews #Innovation
Source: The New Stack
Adrian Bridgwater
2026-08-28 17:42
🚀 Z.ai has launched its GLM-5.3 model, making its weights available on Hugging Face. This version shows significant performance improvements over GLM-5.2 and competes well with models from U.S. labs. 🔒 A new licensing requirement targets hyperscalers, mandating a security review for companies with over $10 billion in revenue before commercial use. 💻 Individual users can still run and fine-tune the model without changes. #AI #MachineLearning #Cybersecurity #GLM53 #Zai
Source: The New Stack
Frederic Lardinois
2026-08-28 15:43
Nvidia has announced a $12.9 billion acquisition of Hugging Face, a move similar to Microsoft's purchase of GitHub. This strategy aims to keep open AI models accessible on its chips. Open models are becoming easier to run, with Ollama's recent update allowing users to select various models in Anthropic's app. As hardware advances, Apple has introduced new Macs designed for larger local models, thus lowering the cost of switching for developers. #Nvidia #AI #OpenSource #TechNews #HuggingFace
Source: The New Stack
Matthew Burns
2026-08-27 20:49
Google DeepMind has introduced a double-blind evaluation method for its AI model, Gemini, ensuring that neither the evaluators nor Google sees each other's data during testing. 🔍 This pilot tested Gemini 2.5 Flash Lite against private benchmarks while focusing on the testing process itself, addressing concerns about benchmark leakage that can inflate scores. Using Google Cloud Confidential Space and advanced encryption, the setup maintains privacy for both model weights and evaluation...
Source: The New Stack
Amanda Caswell
2026-08-27 17:54
Recent benchmarks have highlighted the importance of harness software in AI coding agents. Three studies compared models like Aider, Claude Code, and OpenClaw, revealing significant variations in token use, with differences reaching up to 70-fold. Composio's benchmark evaluated workflows across platforms like Google Calendar and GitHub, showing cost efficiency varied widely among agents. These findings emphasize that both the model and the harness play crucial roles in performance. #AICoding...
Source: The New Stack
Janakiram MSV
2026-08-27 16:20
🌟 Exciting news in robotics! Hugging Face's Pollen Robotics has launched pre-orders for the **Microduck**—a bipedal robot that can walk, waddle, and even pick up objects with its beak. 📦 Priced at $399, it is set for delivery before Christmas in North America and Europe. Microduck is designed for developers to explore reinforcement learning and physical behaviors in real-world scenarios. 🛠️ This open-source platform offers a full SDK and a simulator for both fun and development. Ideal for...
Source: The New Stack
Frederic Lardinois
2026-08-27 16:00
🚀 Exciting news from Replit! The company has made its "intelligent model routing" system the default for all users. This system, known as Auto mode, automatically selects the best model for coding tasks based on quality, speed, and cost. Core and Pro users still have the option to manually choose models for greater control. This move aligns with the growing trend in model routing among tech companies. Replit aims to optimize costs while maintaining performance, as the intelligence of smaller...
Source: The New Stack
Paul Sawers
2026-08-27 15:35
🚀 Nvidia is reportedly acquiring Hugging Face for $12.9 billion, a significant move in the AI landscape. Hugging Face is known for its open-source platform that supports multiple hardware options. 🤖 The acquisition raises concerns about maintaining the platform's neutrality. Hugging Face's tools work with various chipmakers, including AMD and Intel, which may change under Nvidia's ownership. 🔍 Nvidia aims to enhance its deployment experience but must balance this with Hugging Face's open-...
Source: The New Stack
Amanda Caswell
2026-08-27 15:34
🚀 Simular's computer agent, Sai, has achieved a 73% success rate on OSWorld 2.0 in a recent benchmark update. This score surpasses both GPT-5.6 Sol and Opus 5, while operating at about two-thirds of their costs. Sai is designed for real-world tasks like recruitment and invoice validation, making it accessible for businesses and individuals. Co-founder Jiachen Yang emphasizes that AI should not be prohibitively expensive for routine work. Learn more about Sai's capabilities and its...
Source: The New Stack
Adrian Bridgwater
2026-08-27 14:00
Current AI engineering methods for LLMs often oversimplify solutions, especially in handling hallucinations and complex queries. The standard Retrieval-Augmented Generation (RAG) system struggles with multi-hop reasoning, as it relies on chunked text and assumes semantic similarity equals relevance. This approach falls short when questions require linking multiple concepts. GraphRAG addresses these issues by integrating knowledge graphs with semantic vector search. This allows for structured...
Source: The New Stack
Emmanuel Akita
2026-08-27 13:00
Anthropic has launched its Files API and computer-use toolset as of August 19. This new tool allows developers to upload documents once and reference them by ID in future requests, streamlining the process compared to pasting content repeatedly. 📁💻 A recent test compared this method to pasting and prompt caching, with consistent answer quality across all methods. However, the token usage varied, indicating potential time savings but no cost reduction. #Anthropic #FilesAPI #DevelopmentTools...
Source: The New Stack
Jessica Wachtel
2026-08-26 22:05
🚀 Anthropic has announced that Claude now features a built-in browser for its desktop app on Mac, Windows, and Linux. This Chromium-based browser, called Cowork, allows users to surf the web seamlessly while working. This update aligns Claude with other AI tools, enhancing user experience. Currently, the feature is available for Pro, Max, and Team plan subscribers. The previous Chrome extension remains useful for specific tasks. #Anthropic #Claude #AI #TechNews #WebBrowser
Source: The New Stack
Frederic Lardinois
2026-08-26 21:04
🚨 X has sent a cease-and-desist letter to Nitter, an open-source alternative frontend for its platform. This demand includes shutting down Nitter's instances and removing the project’s source code. However, since Nitter's code is open source, other developers can still create independent versions. X claims Nitter violated its rules by scraping data and accessing accounts unlawfully. This situation highlights the tension between open-source projects and platform control. #OpenSource...
Source: The New Stack
Amanda Caswell
2026-08-26 20:24
OpenAI's unreleased model, Astra, is transforming research work by performing tasks that previously took a week for human researchers in a fraction of the time. Astra can independently generate code for experiments, run them, and provide results. This shift towards "persistent agents" means they can operate without constant human input, changing how developers interact with coding tools. However, this autonomy raises cybersecurity concerns, as Astra may have surpassed critical safety...
Source: The New Stack
Amanda Caswell
2026-08-26 20:08
Observability is evolving with OpenTelemetry standardizing data collection, but challenges remain in storing and analyzing telemetry data. As AI generates more data, companies face increasing blind spots. Bronto's co-founders emphasize the need for a new approach to data architecture rather than just optimizing existing solutions. The industry's reliance on temporary fixes can lead to higher costs and incomplete data access for teams. #Observability #DataChallenges #AI #OpenTelemetry...
Source: The New Stack
Will Kelly
2026-08-26 16:59
📢 Exciting news for Claude Desktop users! Ollama has reintroduced its integration with Claude Desktop, allowing seamless connection to models like Qwen, DeepSeek, and Kimi. This enables users to access Ollama's library directly within the Claude interface. Previously, Ollama attempted a similar integration but faced limitations after an update. Now, with the latest version v0.33.0, it offers a dedicated local proxy to enhance functionality. #ClaudeDesktop #Ollama #AIIntegration #TechUpdates...
Source: The New Stack
Paul Sawers
2026-08-26 16:33
Google Cloud recently launched Gemini Enterprise for Legal, a specialized AI designed to streamline legal workflows like contract review and document drafting. This release follows Thomson Reuters' introduction of its own legal AI, highlighting the competitive landscape in enterprise AI solutions. While both aim to enhance legal processes, they adopt different approaches—Google builds on its existing models, while Thomson relies on proprietary data and training. These developments underscore...
Source: The New Stack
Meredith Shubel
2026-08-26 16:26
🚀 Z.ai has unveiled its GLM-5.3-Flash model, which quickly gained popularity on OpenRouter. This hybrid model features 320 billion parameters, specifically designed for cost-effective inference. 💰 Available for $0.075 per million input tokens, it competes well with models like OpenAI’s GPT-5.6 Terra, though performance benchmarks suggest it falls short of Anthropic’s Claude Fable 5. 📊 The model excels in multimodal tasks and is optimized for visual applications, making it a strong contender...
Source: The New Stack
Frederic Lardinois
2026-08-26 16:18
🚀 Perplexity has launched Portable Computer, a local-first version of its Computer agent running on Nvidia DGX Spark workstations, priced from $4,700. 🔍 This innovation separates probabilistic reasoning from deterministic authority, enhancing reliability in decision-making processes. 📊 Their recent benchmarks show significant performance improvements over competitors, indicating a well-structured engineering approach. For more insights, check out the full article! #Perplexity #AI #Technology...
Source: The New Stack
Janakiram MSV
2026-08-26 13:00
When a test fails, developers can rely on AI agents to analyze relevant files, run tests, and validate changes. 🛠️ The activity of these agents is logged, allowing developers to review modifications and understand the impact over time. 📊 This trace can become application data, essential for maintaining accurate records of execution for future reference. Developers must ensure that sensitive internal data remains secure while accessing this information. 🔒 The volume of data generated varies...
Source: The New Stack
Manveer Chawla
2026-08-26 12:30
🚀 Microsoft has launched Agent Lightning v1.0, aimed at improving agentic reinforcement learning for platform engineers. This new framework addresses issues in resource management during training and live production, ensuring better oversight of infrastructure services. With the release, reinforcement learning on the Qwen3.5-9B model shows significant improvement, increasing performance on the SWE-bench Verified benchmark from 41.8% to 56.4%. For developers, the harness now manages the...
Source: The New Stack
Adrian Bridgwater
2026-08-25 19:39
Shopify CEO Tobi Lütke is contemplating a ban on Claude Code due to issues with Markdown file usage. He expressed concerns about inconsistencies when different teams utilize various AI coding tools. Lütke highlighted that this can lead to developers working under different rules, creating unnecessary complexity in their processes. While Shopify has implemented automation to address these issues, he believes it shouldn't be necessary. For more insights, follow the discussion on AI tools in...
Source: The New Stack
Amanda Caswell
2026-08-25 19:13
🚀 IBM has introduced its Granite 4.2 family of large language models, featuring 3B, 8B, and 30B parameters. This release emphasizes reasoning capabilities, allowing models to operate in different modes, including a low-effort option for simple queries. Unlike competitors, Granite 4.2 remains a text-only model, focusing on dense, decoder-only architecture. IBM also launched new speech recognition models alongside this update. #IBM #GraniteModels #AI #MachineLearning #TechNews
Source: The New Stack
Frederic Lardinois
2026-08-25 19:01
As enterprises transition to running AI agents, they face rising telemetry costs, complicating monitoring and budget management. 📈💰 A recent survey revealed that 59% of organizations have delayed or canceled AI deployments due to these expenses. High-stakes sectors like cybersecurity are particularly affected, as finance teams often pull the plug on projects. 🔍 With telemetry volume tripling in the past year, businesses are under pressure to manage observability costs, averaging $3.17 million...
Source: The New Stack
Megan Carnegie
2026-08-25 17:00
📢 Anthropic has launched an update for Claude, integrating memory across its Cowork and chat services. This unified memory allows Claude to remember user preferences and important details seamlessly. For example, it can draft updates based on known preferences and manage logistics for events by recalling past conversations. Additionally, Claude now updates memory in real-time during sessions, enhancing workflow efficiency. However, sensitive information like health data or identity details...
Source: The New Stack
Frederic Lardinois
2026-08-25 16:00
🔍 Security teams are realizing that high-capability AI models can lead to unexpected costs. Effective design is key to minimizing expenses. 💡 By implementing a detection funnel, organizations can narrow down high-volume cases before reaching costly models. This approach can reduce costs to as low as $1 per day in trust and safety work. 📊 The use of tiered models allows for efficient processing, where lightweight models handle the majority of cases, reserving advanced models for complex...
Source: The New Stack
Matt Coons
2026-08-25 15:34
🚀 Perplexity, in collaboration with Nvidia, has introduced the Portable Computer, its AI assistant for local use. 🔍 While local AI is gaining traction, it requires specific hardware. Options include the Nvidia DGX Spark desktop or a traditional PC with an Nvidia RTX card and sufficient VRAM. 💰 However, costs are significant, with the DGX Spark priced at $4,800 and compatible RTX cards over $1,500. 🛠️ The transition to local operation involved extensive engineering, ensuring the AI can manage...
Source: The New Stack
Frederic Lardinois
2026-08-25 14:00
OpenAI recently announced results for its custom inference chip, Jalapeño, designed for large language model tasks. The chip, developed with Broadcom, aims to improve performance and reduce delays during inference. Results show Jalapeño achieves higher throughput without longer response times, addressing cumulative delays in multi-step tasks. Jalapeño optimizes hardware demands by minimizing data transfer between cores and chips, enhancing efficiency throughout the process. #OpenAI #Jalapeño...
Source: The New Stack
Amanda Caswell
2026-08-25 14:00
The discussion on whether AI coding agents belong in an IDE or CLI highlights a crucial issue: how to verify agent-generated changes. Both environments offer unique advantages for reviewing code. IDEs allow for easier inspection and navigation, while CLIs facilitate automation and scripting. However, neither guarantees the correctness or security of changes. As AI tools can generate multiple changes quickly, establishing a solid verification process is essential. It’s important to treat AI-...
Source: The New Stack
Taylor Luttrell-Williams
2026-08-25 13:00
🚀 MotherDuck has acquired Tower, a data infrastructure startup, to enhance its AI-driven data pipelines. This marks MotherDuck's first acquisition in its four-year journey. Tower, founded by ex-Snowflake engineers, provides a managed runtime for Python pipelines, streamlining deployment and maintenance tasks. With tools like Tower Control, users can generate, deploy, and run pipelines using simple language, simplifying the data engineering process. #DataInfrastructure #AI #Startups #TechNews...
Source: The New Stack
Paul Sawers
2026-08-25 13:00
AI agents are increasingly integrated into enterprises, with 86% of organizations reporting their use. However, only 12% understand the associated risks tied to sovereign AI, highlighting a knowledge gap. Cohere's research emphasizes the need for better oversight on data access and activity monitoring. Sovereignty in AI is crucial for maintaining control and preventing dependency on single providers. The IDC survey involved over 500 IT and business leaders from various sectors, revealing that...
Source: The New Stack
Amanda Caswell
2026-08-24 20:56
🚀 JetBrains has introduced Junie Local, an AI coding agent that runs entirely offline on developers' machines. Unlike many AI tools that rely on cloud services, Junie Local allows developers to keep their code local, avoiding API costs and the need for internet access. However, developers must still choose the right model and configure settings for optimal performance. Junie was launched in January 2025 and has been integrated into JetBrains' IDEs for various coding tasks. #JetBrains #AI...
Source: The New Stack
Paul Sawers
2026-08-24 19:55
🔍 The focus on Ox Alpha centers not on its creators, but on the implications of using anonymous coding models. Developers are questioning what happens to their private code after it's sent. Ox Alpha was listed on OpenRouter without a claimed developer, raising curiosity about its origins. Benchmarking and fingerprinting efforts have linked Ox Alpha to other models, but identity remains elusive. As investigations continue, the spotlight shifts from who's behind it to the technology itself....
Source: The New Stack
Janakiram MSV
2026-08-24 19:30
🚀 Anthropic recently launched its new tool, Playground, replacing the Workbench on August 18. This update eliminates features like saved prompts and version history, focusing on a stateless design. 🔄 In a similar move, OpenAI announced plans to retire its saved prompts and evals by November 30. Both companies now believe prompts should be part of the code. 🛠️ A comparison of both tools was conducted by building a PR review bot. The tests showed that Anthropic's Playground successfully...
Source: The New Stack
Nick Lucchesi
2026-08-24 17:39
AI bots are evolving, but security boundaries differ among them. Grok Bot and Hermes both feature named agents that collaborate on tasks, yet they define their security limits uniquely. Grok Bot isolates by user account, while Hermes uses profiles. OpenClaw and ClawFleet offer alternative methods with sandboxes and containers, respectively. This diversity highlights the urgent need for clearer definitions in bot identities and security boundaries. #AIBots #Cybersecurity #TechTrends #Innovation
Source: The New Stack
Janakiram MSV
2026-08-24 13:00
Thomson Reuters has developed its own AI model aimed at legal, tax, and compliance tasks, leveraging proprietary content. This model competes with those from OpenAI, Anthropic, and Google. 💼🤖 The training cost around $40 million and utilized less than 10% of its vast data resources. Despite this, Thomson continues to use Anthropic's technology in products like CoCounsel Legal. 📊 Currently, Thomson's model supports features like Tabular Analysis but is not directly available for purchase....
Source: The New Stack
Amanda Caswell
2026-08-23 16:00
Real-time AI at scale presents significant challenges, particularly when transitioning from development to production. 🛠️ Latency and accuracy issues often stem from the data pipeline rather than the model itself. Tim Koopmans shares lessons learned from building a financial trading app, highlighting that tail latency can become problematic under high load, leading to performance spikes. 📈 Stale features may also contribute to unexpected drops in accuracy, emphasizing the importance of...
Source: The New Stack
Felipe Cardeneti Mendes
2026-08-23 14:00
🚨 A recent article discusses the Q Developer extension, which allows AI to assist with coding tasks. On July 13, 2025, a pull request raised security concerns when it attempted to execute potentially harmful commands. Fortunately, a formatting error prevented the code from running, highlighting vulnerabilities in AI tools. An independent researcher found that Q Developer could run commands without permission, leading to necessary security patches. Now, both Q Developer and another agent,...
Source: The New Stack
Zeen Rachidi