0-Click Exploit Chain for Pixel 10 Disclosed
Pixel 10 零点击漏洞链曝光 ⭐️ 9.0/10

Google Project Zero disclosed a 0-click exploit chain for the Pixel 10 that chains two vulnerabilities (a Dolby audio bug and a VPU kernel bug) to achieve root-level compromise without any user interaction. This exploit chain highlights the increased attack surface introduced by AI-driven message processing, which pre-decodes media and creates new 0-click vectors, and underscores the need for faster patching and more secure system architecture. The chain exploits CVE-2025-54957 (Dolby) and a VPU kernel bug; the VPU bug was patched in the February 2026 Pixel security update. Google's response time was notably fast, with the bug fixed within 90 days of disclosure.

hackernews · happyhardcore · May 15, 13:39 · Discussion

Background: A 0-click exploit requires no user action, making it highly dangerous. AI-powered features that automatically decode messages (e.g., for search or preview) increase the 0-click attack surface by processing media before the user opens it. Google Project Zero researches such vulnerabilities to improve security across Android.

Discussion: Commenters expressed concern that AI-driven message processing increases attack surface, with one noting 'haven't we learned our lesson on this?' Others praised Google's fast response time for this exploit, while some discussed the apparent increase in published exploits and questioned Apple's jailbreak scene.

Tags: #security, #exploit, #Android, #Project Zero, #vulnerabilities

Erlang/OTP 29.0 Released with Security Fixes
Erlang/OTP 29.0 发布，包含安全修复 ⭐️ 9.0/10

Erlang/OTP 29.0 has been released, introducing several improvements including disabling SSH daemon and SFTP by default, the new io_ansi module, and critical security patches addressing CVEs. This major version release addresses critical security vulnerabilities that require immediate attention, and the new stdlib features like io_ansi enhance Erlang's capability for building CLI applications. Notable changes include the disabling of SSH daemon and SFTP by default for improved security, and the addition of io_ansi module to stdlib for ANSI terminal support. The release also includes CVE fixes dated February to May 2026. (Note: likely a typo, should be 2025 or 2026?)

hackernews · pyinstallwoes · May 15, 23:33 · Discussion

Background: Erlang is a general-purpose concurrent functional programming language and runtime system designed for building massively scalable soft real-time systems with high availability. OTP (Open Telecom Platform) is a set of middleware, libraries, and tools that are an integral part of Erlang's distribution, standardizing creation of reliable fault-tolerant applications.

Discussion: Community comments welcome the improvements, particularly security changes like disabling SSH daemon by default. Users note the need to update production apps immediately due to critical CVEs, and express interest in how new features like io_ansi and records will impact the ecosystem. Some also inquire about WhatsApp's continued use of Erlang.

Tags: #Erlang, #OTP, #Release, #Security, #Programming Language

Nvidia Releases 2.6B Open-Source World Model
英伟达发布 2.6B 开源世界模型 ⭐️ 9.0/10

Nvidia has released a 2.6B parameter open-source world model that generates controllable worlds from a single image, text prompt, or trajectory, and it runs on a single GPU such as an RTX 5090 or H100. This makes advanced world models accessible to a wide range of researchers and developers, especially for robotics and simulation research, democratizing technology that previously required large computational clusters. The model is fully open-source, with code available on GitHub and a paper on arXiv. It is designed to run efficiently on a single consumer or datacenter GPU, lowering the barrier to entry for embodied AI research.

rss · Paul Couvert(@itsPaulAi) · May 15, 21:39

Background: World models are AI systems that build an internal representation of an environment and predict how it evolves in response to actions. They are essential for embodied AI, where AI interacts with the physical world through sensors and actuators. Nvidia's release makes such models more accessible for experimentation and deployment.

Tags: #world model, #Nvidia, #open-source, #AI, #robotics

Mitchellh criticizes blind AI adoption in companies
Mitchellh 批评企业盲目采用 AI ⭐️ 8.0/10

Mitchellh, creator of Vagrant and Consul, posted a tweet denouncing companies for irrationally embracing AI without critical thinking, calling it 'AI psychosis.' This highlights a growing concern about overhyping AI in the tech industry, where companies may adopt AI for appearance rather than value, leading to potential waste and misuse. The tweet garnered strong community engagement (1218 points, 567 comments) and sparked discussion on Hacker News about the dangers of blind AI adoption and outsourcing decision-making to AI.

hackernews · reasonableklout · May 15, 20:26 · Discussion

Background: AI psychosis refers to the irrational adoption of AI tools without critical evaluation, often driven by competitive pressure or hype. Mitchellh is a respected figure in DevOps, and his critique resonates amid widespread AI integration that sometimes leads to poor outcomes.

Discussion: Commenters shared mixed experiences: some reported successful AI use without issues, while others warned against outsourcing thinking to AI. One noted CFOs feeling pressure to catch up on AI adoption, and another described non-engineers making critical infrastructure decisions via AI ('vibe coding'), risking disasters.

Tags: #AI, #tech industry, #critical thinking, #hype, #software engineering

California bill mandates patches or refunds for dead online games
加州法案要求停运网游提供补丁或退款 ⭐️ 8.0/10

A California bill has passed a key hurdle that would require game publishers to release patches or offer refunds when an online game is shut down, aiming to preserve playability or compensate consumers. This legislation could set a precedent for digital preservation and consumer rights in the gaming industry, potentially forcing publishers to consider long-term support or face financial penalties. The bill reportedly exempts games offered solely during a subscription period, which could accelerate the shift to subscription models. Publishers may also be required to give 60 days' notice before shutdown.

hackernews · Lihh27 · May 15, 19:48 · Discussion

Background: Online games often become unplayable after server shutdowns, leaving consumers with no recourse despite having purchased in-game items or subscriptions. Similar bills have been proposed in other states but none have passed. This bill seeks to address the lack of digital ownership and consumer protection in the gaming industry.

Discussion: Commenters are divided: some support open-sourcing server code to let communities run their own servers, while others worry the bill will increase risk and cost for game developers, potentially reducing the number of online games created. A game developer shutting down their own game notes the high ongoing costs of moderation and server maintenance, arguing the bill might just make it more likely that a failing game drags down the entire company.

Tags: #gaming, #digital preservation, #legislation, #consumer protection, #software shutdown

Zulip Forms Nonprofit Foundation as Core Team Joins Anthropic
Zulip 成立非营利基金会，核心团队加入 Anthropic ⭐️ 8.0/10

Zulip announced the formation of an independent nonprofit foundation to oversee the open-source project, as its core team (including creator Tim Abbott) departs to join AI company Anthropic. This transition ensures Zulip's long-term independence and community governance, addressing trust concerns about commercial pressures in open-source projects. It also highlights the ongoing talent migration from open-source to AI companies. The foundation is established with a mission to serve the public good, and the departing team members are donating the company to the foundation. The announcement was made on a Friday afternoon, which some community members noted as an unusual timing that could be intended to minimize attention.

hackernews · boramalper · May 15, 18:37 · Discussion

Background: Zulip is an open-source team chat and collaboration software known for its topic-based conversations. Anthropic is an AI safety company valued at $380 billion as of February 2026. The Zulip core team's move to Anthropic mirrors similar trends in open-source, such as the Bun announcement earlier that week.

Discussion: Community reaction is mixed: some express sadness about the core team leaving but excitement for the foundation's independence (e.g., pig208: 'a bit sad about the departure... this is not that'), while others are skeptical about the Friday timing and compare it to Bun's acquisition (tiffanyh). Many users praise Zulip's interface and wish more projects adopted it (conorbergin).

Tags: #opensource, #foundation, #zulip, #nonprofit, #community

US DOJ seeks identities of 100k car-tinkering app users
美国司法部要求披露 10 万汽车改装应用用户身份 ⭐️ 8.0/10

The U.S. Department of Justice has demanded that Apple and Google reveal the identities of over 100,000 users of a car-tinkering app used to bypass emissions controls. This demand raises significant privacy concerns and questions about government overreach, while also highlighting the risks of centralized app distribution through major platforms. The app in question allows users to disable factory-installed emissions controls, and the DOJ claims it needs the information to identify witnesses for an ongoing investigation into emissions tampering.

hackernews · tencentshill · May 15, 17:28 · Discussion

Background: Defeat devices are illegal modifications that bypass emissions controls, often through ECU remapping. The EPA has been cracking down on such devices, which can significantly increase harmful emissions. This legal demand could set a precedent for future requests targeting users of legally questionable apps.

Discussion: Comments are divided: some argue the government should go after manufacturers rather than users, while others warn of a slippery slope where such precedent could be used to target car modifications like disabling GPS tracking. Some suggest using anonymous app stores like F-Droid to avoid surveillance.

Tags: #privacy, #government surveillance, #emissions, #app store, #legal

Hacker News Celebrates Jason Scott's Digital Preservation Work
黑客新闻社区赞誉 Jason Scott 的数字保存工作 ⭐️ 8.0/10

Jason Scott's ASCII blog and his extensive digitization of historical media have been highly praised by the Hacker News community, with users highlighting his decade-long dedication to preserving digital artifacts. This recognition underscores the importance of individual efforts in digital preservation, especially as physical media decays and knowledge risks being lost. Jason Scott's work ensures that historical software, documents, and cultural artifacts remain accessible to future generations. Jason Scott has digitized over 1,300 tapes from a lifetime collection of magnetic media, and a collection of 13,000 manuals was uploaded to the Internet Archive over ten years, averaging about 3.5 manuals per day.

hackernews · bookofjoe · May 15, 14:02 · Discussion

Background: Jason Scott is a digital archivist known for his work with the Internet Archive and the Archive Team. He runs the blog 'ASCII' at textfiles.com, where he documents his digitization projects, often focusing on retrocomputing and preserving historically significant but endangered data.

Discussion: Commenters expressed deep gratitude and admiration for Jason Scott's prolific output, with one user noting he digitized over 1,300 tapes from their personal collection. Another highlighted the remarkable pace of uploading 13,000 manuals over a decade. The overall sentiment was highly positive, celebrating his dedication and impact.

Tags: #digital preservation, #Jason Scott, #Internet Archive, #retrocomputing, #archiving

Satirical article criticizes npm's recurring supply chain attacks
讽刺文章批评 npm 反复出现的供应链攻击 ⭐️ 8.0/10

A satirical blog post titled "No way to prevent this," says only package manager where this regularly happens critiques npm's inability to prevent supply chain attacks, sparking community debate on mitigation strategies. npm is the largest JavaScript package registry, and recurring supply chain attacks (such as those affecting axios and tanstack) compromise countless applications. The debate highlights the urgent need for better security defaults and practices in the open-source ecosystem. Community members proposed solutions including package cooldowns (ignoring packages younger than N days), reproducible builds with signed attestations, and disabling postinstall scripts by default. The article and comments reflect frustration with npm's slow adoption of security improvements.

hackernews · alligatorplum · May 16, 00:36 · Discussion

Background: A supply chain attack in software occurs when malicious code is introduced into a legitimate package, which then propagates to downstream users. npm, the default package manager for Node.js, has been a frequent target due to its massive ecosystem and permissive defaults like automatic execution of postinstall scripts. Recent high-profile compromises include the axios package and others, affecting millions of downloads.

Discussion: Commenters generally agreed that npm's security is inadequate, with proposals ranging from simple cooldowns (eranation) to more robust measures like reproducible builds and signed attestations (xiaosong001). Some argued for completely disabling postinstall scripts (827a), while others noted the difficulty of enforcing safe configurations across teams (joeblubaugh). The sentiment was a mix of frustration and practical suggestions.

Tags: #npm, #supply chain attack, #package management, #security, #open-source

ChatGPT Now Links Bank Accounts for US Pro Users
ChatGPT 现可绑定美国 Pro 用户银行账户 ⭐️ 8.0/10

OpenAI has released a new feature for US Pro users that allows them to link financial accounts—such as bank, investment, and credit card accounts—directly within ChatGPT, using Plaid to connect over 12,000 institutions. Users can view a unified asset dashboard, ask questions based on real-time data, and perform actions like applying for credit cards or scheduling tax appointments through a partnership with Intuit. This marks a significant step for OpenAI as ChatGPT evolves from offering advice to executing financial actions, potentially disrupting personal finance apps like Monarch, Copilot, and YNAB. While currently limited to US Pro users, it signals a broader trend of AI agents handling sensitive real-world tasks, with implications for privacy and competition in fintech. Linked accounts show balances, transactions, investments, and liabilities but never full account numbers, and ChatGPT cannot move money. Data sync is deletable within 30 days, and a separate 'financial memories' feature stores user-provided financial context. The feature uses GPT-5.5 Thinking by default, with GPT-5.5 Pro available for Pro users, scoring 79 and 82.5 respectively in internal evaluations by financial professionals.

rss · 宝玉(@dotey) · May 15, 19:10

Background: Plaid is a financial data aggregation platform that allows users to securely connect their bank accounts to third-party apps. Intuit is the company behind TurboTax and QuickBooks, and previously owned the popular budgeting app Mint. This partnership with Intuit enables ChatGPT to not only provide advice but also execute actions like submitting credit card applications or scheduling tax consultations within the same conversation.

Tags: #OpenAI, #ChatGPT, #金融科技, #AI代理, #产品更新

OpenAI Merges ChatGPT and Codex into Unified Team
OpenAI 将 ChatGPT 与 Codex 合并为统一团队 ⭐️ 8.0/10

OpenAI announced an internal reorganization that merges ChatGPT, Codex, and the API business into a single core product team led by president Greg Brockman. This merger signals OpenAI's strategic shift toward product integration ahead of a potential IPO, and positions it to better compete with rivals like Anthropic's Claude and Google in both developer and consumer markets. Codex head Thibault Sottiaux has been promoted to lead the combined product and platform division, while ChatGPT leader Nick Turley shifts focus to enterprise product redesign. Brockman's internal memo cited the need for a unified experience as Codex increasingly powers OpenAI's consumer offerings.

rss · 宝玉(@dotey) · May 15, 18:41

Background: OpenAI Codex is an AI coding agent released in April 2025 that can write code, fix bugs, and operate within a secure container. ChatGPT is a conversational AI chatbot that reached over 900 million weekly active users. The reorganization comes as OpenAI prepares for a potential IPO and faces competitive pressure from Anthropic's Claude and Google's products.

Tags: #OpenAI, #ChatGPT, #Codex, #product strategy, #reorganization

Arena's Data Pipeline for Tagging Millions of Weekly Votes
Arena 的千万级周投票数据标注管道 ⭐️ 8.0/10

Arena researchers Guanglei Song and I-Hung Hsu presented the data pipeline that uses Databricks, Spark, and a pluggable tagger framework with LLMs to categorize millions of weekly evaluations across multiple arenas. This metadata layer enables researchers to use Arena data for more than just leaderboard rankings, providing structured categories for deeper analysis. The pipeline handles flaky LLM APIs with dynamic concurrency control, supports adding new taggers without rebuilding the system, and implements cost control via filtering, idempotency, and model selection. It also chunks long messages and backfills history alongside the live stream.

rss · Arena.ai(@lmarena_ai) · May 15, 15:57

Background: Arena (formerly Chatbot Arena) is a public platform where users compare anonymous LLMs and vote on better responses, generating millions of votes weekly. The platform uses Elo ratings to rank models based on crowd-sourced pairwise comparisons. The new pipeline adds a tagging system that categorizes each evaluation into specific topics or capabilities.

Tags: #data pipeline, #LLM, #Arena, #tagging system, #Databricks

New Open-Weight Models Top Document Parsing Leaderboard
新型开放权重模型登顶文档解析排行榜 ⭐️ 8.0/10

INF released two open-weight models, Infinity-Parser2-Pro (35B) and Infinity-Parser2-Flash (2B), that top the ParseBench leaderboard for document understanding. The models use an expanded synthetic data engine with over 5 million diverse parsing samples and a novel Joint RL algorithm that co-optimizes multiple complex tasks. These models advance the state of the art in document parsing for AI agents, offering open-weight alternatives that can reliably structure enterprise documents. The novel Joint RL training approach demonstrates a promising direction for handling diverse parsing tasks in a unified manner. ParseBench is an open benchmark with about 2,000 human-verified enterprise document pages and over 167,000 deterministic test rules across five dimensions: tables, charts, content faithfulness, semantic formatting, and visual structure. The Joint RL algorithm co-optimizes tasks including document parsing, element parsing, and chart parsing.

rss · Jerry Liu(@jerryjliu0) · May 15, 21:51

Background: Document parsing converts scanned PDFs into structured, machine-readable formats that AI agents can act on. Traditional methods rely on pipeline-based OCR and rule-based extraction, but recent advances use large vision-language models (LVLMs) trained end-to-end. Reinforcement learning is emerging as a technique to improve layout awareness and parsing accuracy beyond supervised fine-tuning.

Tags: #Document Understanding, #Open-Weight Models, #AI/ML, #NLP, #Benchmark

LangChain Unveils Major Product Suite Update at Interrupt 2026
LangChain 在 Interrupt 2026 发布重大产品套件更新 ⭐️ 8.0/10

LangChain announced multiple new products including LangSmith Engine for automated issue detection and fix generation, SmithDB for high-performance agent observability, and Managed Deep Agents for production-ready deployment. This comprehensive update significantly streamlines the development workflow for LLM-based agents, reducing debugging time and improving reliability, which could accelerate enterprise adoption of agentic AI systems. LangSmith Engine replaces manual trace analysis by clustering failures and drafting PRs; SmithDB delivers up to 12x faster query performance compared to previous storage; Managed Deep Agents are model-agnostic and open source.

rss · LangChain(@LangChainAI) · May 15, 15:47

Background: LangChain is a leading framework for building applications with large language models (LLMs). LangSmith is its companion platform for debugging, testing, and monitoring LLM applications. The new products address common challenges in developing reliable agentic systems, such as identifying root causes of failures and scaling observability.

Tags: #LangChain, #LLM, #AI tools, #product launch, #LangSmith

xAI Open-Sources X's 'For You' Algorithm Rewritten in Rust
xAI 开源用 Rust 重写的 X 平台“为你推荐”算法 ⭐️ 8.0/10

xAI has open-sourced the recommendation algorithm powering X's 'For You' feed, rewritten in Rust. The system uses a two-layer architecture: Thunder for real-time in-network posts and Phoenix dual-tower model for out-of-network discovery, with ranking via a Transformer model adapted from Grok-1. This open-sourcing provides unprecedented transparency into X's recommendation system, allowing developers and researchers to study and improve it. The use of Rust and the integration of Grok-1 demonstrate a modern, performance-oriented approach to large-scale recommendation. The in-network layer uses Thunder, a real-time in-memory post store that retrieves posts from followed accounts in sub-milliseconds. The out-of-network layer employs the Phoenix dual-tower model for candidate retrieval from the global corpus, and the ranking model is a decoder-only Transformer with 314B parameters (Grok-1) adapted for multi-objective prediction (likes, replies, retweets, clicks).

rss · Geek(@geekbb) · May 15, 14:19

Background: Recommendation algorithms are central to social media platforms, determining which content users see. X's 'For You' feed mixes posts from followed accounts (in-network) and recommended content from unknown accounts (out-of-network). Rust is a systems programming language known for safety and performance. Thunder is a key-value store optimized for speed, Phoenix is a dual-tower embedding model for candidate retrieval, and Grok-1 is xAI's large language model with mixture-of-experts architecture.

Tags: #open source, #recommendation algorithm, #Rust, #xAI, #Grok-1

v0 Gains Browser Use: Autonomously Opens, Uses, and Debugs Built Apps
v0 新增浏览器使用功能：自主打开、使用和调试所构建的应用 ⭐️ 8.0/10

Vercel's AI tool v0 can now autonomously open the web apps it generates, interact with them, critique their designs, debug complex flows, and proactively fix issues, all while sending real-time screenshots to the user. This marks a significant advancement in AI-driven software development, as v0 transitions from a code generator to a self-testing and debugging agent, potentially reducing developer workload and accelerating iteration cycles. The feature uses browser automation to interact with the built app, and the user receives screenshots during the process. It is part of v0's ongoing evolution from UI generation to full-stack application building with deployment capabilities.

rss · v0(@v0) · May 15, 21:39

Background: v0 is an AI-powered application builder by Vercel, known for generating React components with shadcn/ui and Tailwind CSS from natural language prompts. Originally focused on frontend UI generation, it has expanded to full-stack development, including database connections and API integrations, with one-click deployment to Vercel. This new browser use capability adds autonomous testing and debugging, moving beyond static code generation.

Tags: #AI, #browser use, #debugging, #software testing

Anthropic surpasses OpenAI in enterprise adoption rate for first time
Anthropic 首次在企业采用率上超越 OpenAI ⭐️ 8.0/10

According to the latest Ramp AI Index, 34.4% of enterprises now use Anthropic, surpassing OpenAI at 32.3%, marking the first time Anthropic has led in enterprise adoption. Over the past year, Anthropic's adoption quadrupled while OpenAI's grew only 0.3%. This signals a major shift in the enterprise AI market, as Anthropic's safety-focused approach gains traction over OpenAI's broader brand recognition. It could prompt competitors to accelerate feature development and enterprise-specific offerings. The Ramp AI Index tracks AI product adoption among American businesses, with the latest data from March 2026 showing total adoption crossing 50% for the first time. Anthropic's growth rate of 4x year-over-year contrasts starkly with OpenAI's stagnation.

rss · AI Will(@FinanceYF5) · May 15, 08:22

Background: The Ramp AI Index is a metric that measures the percentage of American businesses using AI products and services, based on transaction data from Ramp's platform. As of March 2026, overall business AI adoption surpassed 50%, up from 35% a year earlier. Anthropic is an AI company known for its Claude model and emphasis on safety, while OpenAI is the creator of GPT series and ChatGPT.

Tags: #Anthropic, #OpenAI, #enterprise AI, #AI adoption, #Ramp AI Index

PhyMotion: Structured 3D Motion Reward for Human Video Generation
PhyMotion：为人体视频生成设计的结构化 3D 运动奖励 ⭐️ 8.0/10

PhyMotion introduces a structured 3D motion reward that leverages a physics simulator to evaluate the physical plausibility of generated human motions, improving the realism of human video generation. This approach bridges the gap between generative video models and physical reality, enabling more lifelike and physically consistent human animations, which is crucial for applications in media, gaming, and simulation. The reward is fine-grained and multi-dimensional, assessing motion quality across various aspects of physical feasibility such as balance, contact forces, and joint limits. It is designed to be integrated into existing video generation pipelines.

rss · AK(@_akhaliq) · May 15, 17:37

Background: Physics-grounded human video generation aims to produce videos where human movements adhere to real-world physics, but current models often produce unnatural motions. The PhyMotion reward provides a principled way to guide generation by using a physics simulator to score motion sequences.

Tags: #AI, #Video Generation, #Physics, #3D Motion, #Machine Learning

Anthropic CFO on $100B Compute and Exponential Growth
Anthropic CFO：百亿美金算力赌注与指数增长 ⭐️ 8.0/10

Anthropic CFO Krishna Rao revealed in a podcast that the company has committed over $100 billion to compute infrastructure, driven by exponential revenue growth from $9 billion to $30 billion annualized within four months, and confirmed that scaling laws remain robust with recursive self-improvement where over 90% of code is now written by Claude. This deep dive from a top AI lab's CFO provides rare insight into the strategic thinking behind massive compute investments, exponential business growth, and the belief that scaling laws are far from slowing down, impacting investors, competitors, and the entire AI ecosystem. Key technical details include flexibility across three chip platforms, self-built compilers for bare-metal control, and a compute baseline that never dips below model development needs. The company also noted a net dollar retention rate exceeding 500% and that 9 out of the Fortune 10 are customers.

rss · 跨国串门儿计划 · May 16, 03:22

Background: Anthropic is an AI safety company founded by former OpenAI employees, known for developing the Claude family of large language models. The company has raised billions from investors including Amazon and Google, and its CFO Krishna Rao oversees massive compute procurement and financial strategy. In the podcast, he discusses the concept of 'recursive self-improvement' where models help design better models, and the Jevons paradox where lowering prices actually increases total compute consumption.

Tags: #AI, #Anthropic, #Compute Infrastructure, #Business Strategy, #Scaling

Cloudflare Workflows V2 Delivers Deterministic Execution and 50K Concurrency
Cloudflare Workflows V2 引入确定性执行，支持 5 万并发实例 ⭐️ 8.0/10

Cloudflare launched Workflows V2, a redesigned distributed workflow orchestration system featuring deterministic replayable execution, improved observability, and the ability to handle 50,000 concurrent workflow instances and 2 million queued workflows. This update significantly enhances reliability and scalability for distributed systems and serverless computing, making Cloudflare's platform more viable for complex AI agent coordination, data pipelines, and background processing at scale. Workflows V2 employs deterministic replayable execution to ensure consistent state recovery, and the observability improvements include better tracing and monitoring capabilities.

rss · InfoQ · May 15, 14:04

Background: Workflow orchestration systems coordinate multiple steps in a distributed application, often involving state management and error handling. Deterministic execution ensures that replaying the same inputs always produces the same outputs, which is crucial for reliability in distributed systems. Cloudflare Workflows is a serverless workflow service built on Cloudflare's global network.

Tags: #Cloudflare, #Workflows, #Orchestration, #Distributed Systems, #Serverless

Discord Voice Outage Caused by Hidden Circular Dependency
Discord 语音中断由隐藏的循环依赖引发 ⭐️ 8.0/10

Discord published a postmortem revealing that a hidden circular dependency in its voice infrastructure triggered a cascading failure, causing a major voice outage on March 25, 2026. This incident highlights how subtle architectural issues like circular dependencies can cause significant outages even in well-engineered systems, offering valuable lessons for improving system reliability and resilience. The circular dependency was previously undetected and caused a cascading failure that disrupted voice services across the entire platform. Discord's detailed postmortem provides insight into the detection and resolution process.

rss · InfoQ · May 15, 12:00

Background: A circular dependency occurs when two or more services depend on each other to function, creating a loop that can lead to failures if one component breaks. In complex distributed systems, such dependencies can be hard to detect until they cause an outage. Postmortems like Discord's are valuable for the engineering community to learn from real-world incidents.

Tags: #postmortem, #Discord, #circular-dependency, #outage

BlackFile Vishing Extortion: MFA Bypass via SSO Compromise
BlackFile 语音钓鱼勒索：通过 SSO 绕过 MFA ⭐️ 8.0/10

Google Threat Intelligence Group (GTIG) published a detailed analysis of the UNC6671 threat actor's BlackFile extortion campaign, which uses voice phishing (vishing) and adversary-in-the-middle (AiTM) techniques to compromise single sign-on (SSO) and bypass multi-factor authentication (MFA), targeting Microsoft 365 and Okta environments. This campaign demonstrates that even MFA-protected cloud environments are vulnerable to sophisticated social engineering attacks, highlighting the urgent need for organizations to adopt phishing-resistant MFA and strengthen identity security. UNC6671 targets employees on personal phones, posing as IT staff to direct them to credential-harvesting sites. The group uses AiTM proxies to capture session tokens, then uses Python and PowerShell scripts to exfiltrate data for extortion.

rss · Cloud Blog · May 15, 14:00

Background: Vishing (voice phishing) is a social engineering technique where attackers use phone calls to trick victims into revealing sensitive information. Adversary-in-the-middle (AiTM) attacks intercept communications between users and legitimate services to steal credentials and session tokens, effectively bypassing MFA. The BlackFile campaign targets SaaS and identity platforms like Microsoft 365 and Okta, exploiting SSO trust relationships.

Tags: #cybersecurity, #threat intelligence, #vishing, #extortion, #MFA bypass

Cerebras Plans $60 Billion IPO
Cerebras 计划 600 亿美元 IPO ⭐️ 8.0/10

Cerebras Systems has announced plans for a $60 billion initial public offering (IPO), reflecting the booming demand for specialized AI hardware. This IPO underscores the soaring valuation of AI chip companies and signals strong investor confidence in specialized hardware for AI workloads, challenging NVIDIA's dominance. Cerebras is known for its wafer-scale chips, such as the WSE-3, which is 57 times larger than the largest GPU and contains 4 trillion transistors.

rss · Latent.Space · May 16, 04:36

Background: Cerebras designs massive single-wafer chips for AI training, bypassing the need for multiple GPUs. The WSE-3 powers the CS-3 supercomputer, capable of training models 10x larger than previous generations. This approach contrasts with NVIDIA's multi-GPU systems, offering potential advantages in memory bandwidth and inter-chip communication.

Tags: #AI Hardware, #IPO, #Cerebras, #Semiconductors, #Financial News

Apple-OpenAI partnership fractures, legal action considered
苹果与 OpenAI 合作关系破裂，或引发法律诉讼 ⭐️ 8.0/10

OpenAI is considering legal action against Apple for insufficient promotion of ChatGPT integration, as subscription conversions fall far short of expectations. Apple plans to open Siri to other AI models like Claude and Gemini at WWDC 2026 in iOS 27, further diluting OpenAI's exclusivity. This dispute could set a precedent for AI model integration contracts and impact how major tech companies partner with AI providers. The rift highlights the challenges of balancing user experience with commercial agreements in the rapidly evolving AI landscape. OpenAI has hired external lawyers to explore legal options and may issue a formal breach notice to Apple soon. ChatGPT's entry point in Apple's system is hidden and limited, causing most users to still use the standalone app, while Apple also objects to OpenAI's privacy standards and hardware poaching.

telegram · zaihuapd · May 15, 12:59

Background: Apple partnered with OpenAI to integrate ChatGPT into iOS, aiming to enhance Siri with advanced AI capabilities. However, the integration faced usability issues and disappointing revenue. Meanwhile, competitors like Anthropic's Claude and Google's Gemini are vying for partnerships with device makers.

Tags: #Apple, #OpenAI, #partnership, #legal, #AI

Google bans manipulation of AI search results
谷歌禁止操控 AI 搜索结果 ⭐️ 8.0/10

Google has updated its spam policy to explicitly classify manipulating generative AI search responses as a violation, covering AI Overview and AI Mode. This includes tactics like generating biased “best recommendation” content or embedding prompts to influence AI citations. This policy targets the emerging GEO (Generative Engine Optimization) tactics, signaling that Google treats attempts to manipulate AI search results as seriously as traditional SEO spam. It could reshape how businesses optimize for AI-powered search features. Violations may lead to ranking demotions or complete removal from Google Search results. The policy covers AI Overview and AI Mode, which uses Google's Gemini model for complex queries.

telegram · zaihuapd · May 16, 06:31

Background: Generative Engine Optimization (GEO) is a practice that aims to influence how large language models (LLMs) like Gemini, ChatGPT, and Perplexity present information in search results. Google's AI Mode, introduced in March 2025, provides AI-generated responses to complex queries. This update formalizes Google's stance against manipulating these AI-driven features.

Tags: #Google, #AI search, #spam policy, #SEO, #GEO