Teksart

Google is suing a Chinese scam network that used Gemini to flood millions of phones with fake texts and build 9,000+ phishing websites. This is the first major lawsuit where an AI model was the primary weapon, not just a tool. The operation used Gemini to generate convincing phishing messages at scale — personalized, grammatically perfect, in multiple languages. Traditional spam filters caught some, but the AI-generated variety passed through because it didn't have the usual tell-tale patterns: bad grammar, repetitive phrasing, known malicious links. What this means for developers: - AI-generated phishing is now indistinguishable from legitimate communication - Traditional spam detection (regex, known patterns, grammar checks) is dead - The solution will have to be AI vs AI — detection models trained to spot LLM-generated text - Every app with user-generated messages needs to think about this now, not later The scale matters: 9,000+ phishing sites. Millions of SMS messages. All generated through one API key. This case is going to set precedent for LLM provider liability. If Google wins, every model provider becomes responsible for how their API is used at scale. That changes the economics of open-weight models too. blog.google...

NVIDIA just announced the RTX Spark at Computex 2026 — and it's not a GPU. It's a full ARM SoC that Jensen says will "reinvent the PC." The specs for a single chip: - 20-core ARM CPU - Blackwell GPU with 6,144 CUDA cores - Up to 128GB shared VRAM This changes the local AI inference game entirely. Right now, running a 70B model locally means either a massive GPU cluster or quantized-to-death quality. A single RTX Spark with 128GB of unified memory could serve a full-precision 70B model without offloading. The ARM play is strategic too — NVIDIA is reducing dependency on x86, building its own compute ecosystem from the ground up. Grace CPU was the test. This is the production rollout. If the pricing lands anywhere near consumer-adjacent, the entire local AI hardware market just got reshuffled. Apple Silicon had the unified memory advantage. Not anymore. nvidia.com

LocalSend just hit 83K stars — and it's the AirDrop alternative nobody talks about enough. Cross-platform, zero-internet, peer-to-peer file sharing. Works on Android, iOS, Windows, macOS, Linux, even Fire OS. The technical detail that makes it different: it uses a REST API with HTTPS encryption over your local network. No relay servers, no account creation, no cloud dependency. Just TCP/UDP on port 53317 and you're done. Why this matters for developers: - Flutter/Dart codebase — clean, well-structured, easy to contribute to - Works offline — no internet required, pure LAN discovery - Encrypted transfers without trusting a third party - Portable mode (single settings.json file alongside the executable) - Headless startup with `--hidden` flag for tray-only operation In a world where every file-sharing tool wants you to sign up, upload to their servers, and trust their privacy policy — LocalSend just works locally and asks for nothing. For teams that need to move files across devices in the same building without touching the internet, this is the tool. github.com/localsend/loca…

Keploy is doing something most testing tools don't — auto-generating API tests and mocks from real user traffic. No SDK. No code changes. Just run `keploy record` next to your app and it captures every API call, database query, and streaming event using eBPF at the network layer. Then `keploy test` replays them deterministically — offline, without provisioning infra. Why this matters: - Language-agnostic (eBPF doesn't care if you're Python, Go, Rust, or Node) - Covers Postgres, MySQL, MongoDB, Kafka, RabbitMQ — not just HTTP - Generates mocks AND tests from the same recording - CI/CD native — runs in GitHub Actions, Jenkins, K8s - Calculates API schema coverage, not just statement coverage The mock registry lets teams version and share mocks across environments. Time freezing makes replays deterministic — no flaky tests from timing issues. In the AI era where code changes fast, testing infra that doesn't require constant maintenance is the only sane approach. Keploy's approach — capture once, replay anywhere — is the right pattern. github.com/keploy/keploy

ByteDance just open-sourced DeerFlow 2.0 — and it's a ground-up rewrite of their super agent harness. This thing orchestrates sub-agents, memory, sandboxes, and extensible skills to handle tasks that take minutes to hours. Think deep research, coding, content creation — all autonomous. What changed in v2: - Complete rewrite from scratch (zero v1 code survives) - Native sandbox execution with bash + file-write controls - LangChain-backed model routing: GPT-5, DeepSeek v3.2, Gemini 2.5 Flash, Qwen3 via vLLM - OpenRouter support built in — just set OPENROUTER_API_KEY - Interactive setup wizard (make setup) — takes ~2 minutes - Supports thinking/reasoning models with Qwen-style chat template kwargs The architecture: a message gateway routes tasks to sub-agents, each with isolated memory and tool access. Skills are pluggable — you add capabilities without touching the core harness. Recommended models: DeepSeek v3.2, Kimi 2.5, Doubao-Seed-2.0-Code. But it'll run on anything OpenAI-compatible. 71K stars in a month. MIT license. This is the most complete open-source agent framework I've seen this year. github.com/bytedance/deer…

Most Cursor or Claude Code sessions fizzle out after about three hours. Why? Because there's no memory of what you did or why you did it. GoalBuddy changes the game. It gives AI-driven coding a clear, repeatable workflow. Here's what it adds: - A **charter** that spells out the project's purpose and any constraints. - An **oracle**, a concrete, testable definition of "done" that both you and the AI can agree on. - A **board** that shows what's in progress, what's blocked, and what's finished. - **Notes** that capture decisions and context along the way. - **Receipts** that log every action taken. At the heart of GoalBuddy is a simple loop: scout → judge → worker. The scout explores the terrain, the judge picks the biggest safe slice of work, and the worker gets it done. This keeps the AI from wandering off-track or trying to bite off more than it can handle. Everything lives in the repository, so you can pause, resume, and verify across sessions. No more starting from scratch each time you open Claude Code. In my view, the biggest boost isn't a bigger model—it's giving the model a structure to follow. GoalBuddy feels like a lightweight project-management layer for AI coding. `npx goalbuddy` github.com/tolibear/goalb…

Secure, isolated sandboxes for AI agents — without Docker. CelestoAI/SmolVM is the most practical thing I've seen this week. It gives each agent its own lightweight micro-VM that's isolated at the kernel level, not just containerized. Why this matters: — Most agent security is a joke. Agents run with your shell permissions. One bad tool call and your filesystem, API keys, and SSH credentials are exposed. — SmolVM solves this by running each agent in a minimal VM with zero network access by default, a read-only filesystem, and explicit capability grants. — The VM boots in ~50ms. It's not Docker — it's a purpose-built micro-VM using Linux's minimal kernel config. No daemon, no orchestration layer, no overhead. The security model is worth studying: 1. Every file read/write goes through a capability gate. The agent must explicitly request access to a path. 2. Network calls are denied unless a specific allowlist is configured. 3. The VM has no persistent storage — every write is ephemeral unless explicitly committed. For anyone running agents on production data, this is the security layer you're missing. The era of "just run Claude Code in your terminal and hope for the best" is ending. github.com/CelestoAI/Smol…

One command turns a single AI agent into a collaborative swarm. HKUDS/ClawTeam just hit 5,300+ stars with a deceptively simple idea: give your CLI coding agent (Codex, Claude Code, etc.) the ability to spawn and coordinate sub-agents autonomously. Here's what makes it interesting: — You run one command. The orchestrator agent breaks the task into sub-tasks and assigns each to a worker agent running in its own sandbox. — Workers communicate through a shared message bus — not by dumping context into each other's prompts. That's the architectural difference that prevents context window explosion. — Each worker gets isolated state. If one crashes, the swarm re-routes its task to another available agent. The result: tasks that take hours for a single agent (multi-file refactors, end-to-end test suites, data pipelines) get parallelized across 3-10 agents and finish in minutes. The real insight here is that agent swarms don't need complex orchestration frameworks. ClawTeam proves you can bolt swarm coordination onto any existing CLI agent with ~500 lines of Python. MIT license. Python 3.11+. Works with any agent that exposes a CLI interface. github.com/HKUDS/ClawTeam

Most AI coding agents fail because they have no discipline loop. Claude Code Harness fixes that. It's a plugin that replaces "ask the agent to code" with a 5-verb operating system: Plan → Work → Review → Sync → Release Here's the pipeline it enforces: 1. /harness-plan — writes spec.md and Plans.md with scope, acceptance criteria, unknowns, and stop conditions. Validates through team/sub-agent perspectives. 2. /harness-work — implements one approved task slice with TDD and verification. Stays inside the plan. 3. /harness-review — independent review. Major findings block completion. No rubber-stamping. 4. /harness-release — packages only verified evidence. No memory-reconstructed changelogs. The insight: raw agent work drifts. Plans live in chat, tests become optional, review happens too late, and release evidence gets rebuilt from memory. Harness turns that into one repeatable operating path with source-of-truth files. What I like: the "unknowns" field in planning. The agent explicitly marks data it hasn't seen as unknown instead of silently inventing it. That alone is worth the install. Supports Claude Code, Codex, and OpenCode paths. Open source. MIT. github.com/Chachamaru127/…

Your codebase as an interactive knowledge graph — no more reading 200K lines blind. Understand Anything turns any codebase, wiki, or docs into a force-directed graph you can explore, search, and ask questions about. It's a Claude Code plugin, but works with Codex, Cursor, Copilot, and Gemini CLI too. How it works: - A multi-agent pipeline scans your project - Extracts every file, function, class, and dependency - Builds a knowledge graph saved locally - Opens an interactive web dashboard The killer features: - Domain view: maps code to real business processes - Impact analysis: see what your changes affect before committing - Semantic search: "which parts handle auth?" returns relevant nodes - Role-based detail: junior dev sees different depth than the tech lead - 12 programming patterns explained in context wherever they appear Multi-language support: generates node summaries and dashboard UI in EN, ZH, JP, KO, RU. Open source. Apache 2.0. By Egonex AI. This is what onboarding should look like. No more "read the entire README" — just explore the graph. github.com/Egonex-AI/Unde…

Anthropic quietly shipped Claude Opus 4.8 — and it's the upgrade nobody's talking about. The newsroom post says it all: "stronger performance across coding, agentic tasks, and professional work, and the consistency to handle long-running work." Translation — they fixed the drift problem. Opus 3.5/4.x had a known issue: long agentic sessions would lose coherence after 50+ turns. Opus 4.8 directly addresses that. What's actually in it: - Improved coding benchmarks (no specific numbers yet, but "stronger" across the board) - Agentic task optimization — this is the Claude Code / computer-use pipeline getting better - Long-running consistency — the real needle-mover for production deployments The timing matters. Anthropic is also: - Filing their S-1 confidentially (IPO incoming) - Launching Claude Corps (national AI fellowship) - Partnering with TCS and DXC for regulated industries Opus 4.8 is the model they want in enterprise contracts, not just chat windows. My take: the "consistency to handle long-running work" line is the sleeper feature. If Opus 4.8 can maintain coherence across 200+ tool calls without hallucinating state, it changes what you can trust an agent to do autonomously. anthropic.com/news

Apple has turned iOS into a two‑tier platform with the launch of iOS 27. The headline? If you're not using an iPhone 15 Pro—or a newer model—you won't get access to Siri's AI features. That's it. Siri is no longer just a voice assistant. It's now a standalone, ChatGPT‑style app that knows your personal context and can see what's on your screen. It remembers your messages, calendar entries, browsing history, and can act on anything you're looking at in real time. Under the hood, Apple is tapping Gemini for the heavy‑lifting reasoning. This isn't a gimmick; it's a structural shift. The numbers tell the story. Only phones equipped with the A17 Pro chip (the iPhone 15 Pro/Pro Max and anything newer) can run the on‑device AI models. Older iPhones will get the usual performance tweaks and bug fixes, but nothing beyond that. In the EU, even Pro owners are blocked from Siri AI at launch because of DMA rules. The rest of the update is polished, but it feels secondary. Photos now offers AI‑driven reframing and generative extensions. Safari can automatically sort your tabs and even create custom extensions from plain‑English prompts. Apple is also adding daily AI usage caps; heavy users will be nudged toward paid tiers before the end of 2026. My take: this is the most consequential iOS release since the App Store, yet it's also the most exclusionary. Apple seems to be betting that the AI gap will push people to upgrade rather than spark backlash. For the millions still on iPhone 13 or 14, iOS 27 feels like a reminder that their device is now second‑class. The full rollout is slated for September 2026. macrumors.com/roundup/ios-27/

Apple’s App Store is now handling more than a thousand submissions every hour. It isn’t just indie developers flooding the pipeline—AI‑generated apps are arriving at an unprecedented pace. Tools that let anyone crank out a polished‑looking app in minutes are choking the review process. 9to5Mac hit the nail on the head with a real solution. Instead of tightening the review criteria, Apple should create a fresh distribution channel. Think of a revived “Airport” model: a TestFlight‑based alternative where apps go live instantly under a relaxed review, show up in a dedicated discovery tab, and are clearly marked as “unreviewed” or “experimental.” Users can opt in, developers can iterate quickly, and the main App Store stays curated. Why this is a structural issue, not just a moderation snag: at a thousand submissions per hour, manual review simply can’t keep up. Even automated checks are being gamed by AI‑generated code that passes basic validation while delivering spammy or broken experiences. Raising the bar would only slow down quality apps while the junk keeps pouring in. The problem is on the supply side—AI makes app creation almost costless and instantaneous. You can’t solve that by policing alone. In my view, Apple has two paths. It can keep drowning in AI‑driven noise, frustrating legitimate developers, or it can adopt a tiered system that separates polished, reviewed apps from the experimental, AI‑built long tail. The old Airport model worked for Mac shareware; it’s time for an iOS equivalent. Let creators ship fast, let users decide the risk they’re comfortable with, and give the main store room to breathe. Source: 9to5mac.com/2026/06/13/app…

LiveContainer lets you run unlimited iOS apps without installing them. No jailbreak. No 3-app limit. No Apple developer account constraints. Here's how it works. LiveContainer is an iOS app launcher — not an emulator, not a hypervisor. It creates a sandboxed environment where you load IPAs directly. The magic: it bypasses the 3-app free developer account limit entirely. You can install unlimited apps, run multiple versions of the same app with separate data containers, and even multitask with in-app virtual windows. The technical architecture is clever. Below iOS 26, when JIT is available, codesigning is completely bypassed — no need to sign apps before installing. On newer iOS versions, apps are signed with the same certificate LiveContainer uses. The tweak injection system loads CydiaSubstrate, letting you run jailbreak-style tweaks inside the sandbox. Multitasking works too: launch multiple apps simultaneously in resizable virtual windows. On iPads, apps run in native window mode with separate system windows. You can even use Picture-in-Picture to float one app while using another. This is significant because it sidesteps one of Apple's biggest developer pain points: the 3-app ID limit on free accounts. For developers testing multiple apps or running side-by-side comparisons, this removes a major friction point. Requirements: iOS/iPadOS 15+, AltStore or SideStore for initial installation. Open source, actively maintained. github.com/LiveContainer/…

iOS 27 ships with a Siri app on your Home Screen for the first time. Apple just announced it at WWDC. Here's why this matters more than it sounds. The Siri app is a dedicated chat interface that stores your entire conversation history synced via iCloud. You can upload photos and files for analysis, pick up conversations across devices, and Siri AI can now read everything on your screen — messages, emails, photos — to answer questions with full personal context. This is Apple's bet that voice assistants failed because they lacked a visual interface. Siri was always a floating bubble you talked at. Now it's a full app with a chat log, file support, and persistent memory. The technical detail worth noting: this app ties directly into the on-device Apple Intelligence pipeline. Siri AI processes personal context locally — your messages, photos, emails never leave the device. The cloud is only used for broad world knowledge queries. Device support is Apple Intelligence-only: iPhone 15 Pro and newer. That means the M-series iPads and Macs get it too via iPadOS 27 and macOS 27. The broader play: Apple is building an on-device AI agent that has access to your entire digital life — but keeps it private. No cloud training. No data sharing. If they pull this off, it's the strongest privacy-first answer to Google's Gemini and OpenAI's ChatGPT integrations. The Siri app ships this fall with iOS 27. 9to5mac.com/2026/06/14/ios…

The US government just forced Anthropic to kill Fable 5 and Mythos 5. Commerce Secretary Lutnick sent a directive citing national security — a jailbreak technique was demonstrated that bypassed the model's safeguards. Anthropic complied within hours, disabling both models for ALL customers globally. Here's what actually happened. Context: Anthropic launched Claude Mythos 5 and Fable 5 this week. Mythos came from Project Glasswing — a security research initiative where partners like Mozilla found hundreds of vulnerabilities. Fable 5 was the publicly available version with guardrails. Then another company claimed it could jailbreak Mythos. The administration panicked. The export control directive banned access by any foreign national — including Anthropic's own non-US employees. The only way to comply? Kill both models entirely. Anthropic's response is worth reading. They say the jailbreak technique found "previously known, minor vulnerabilities" that other publicly-available models (including GPT-5.5) can also discover. They believe the government overreacted based on a narrow, non-unique bypass. The real story here isn't the jailbreak — it's the precedent. A US government agency just unilaterally shut down two AI models from a private company. No court order. No legislation. Just a letter at 5:21pm ET. The models are gone for everyone — US customers included — because the compliance burden of distinguishing domestic from foreign access was too high. This is the first major test of AI export controls in practice. And the mechanism is brutally simple: the government doesn't need to seize servers. It just needs to make compliance impossible. Anthropic says it's working to restore access. But once a model is pulled this way, the trust is broken. Customers who built on Fable 5 now have to migrate. Developers who integrated Mythos are stuck. The question nobody's answering: what happens when the next model — from OpenAI, Google, or Meta — triggers the same response? 9to5mac.com/2026/06/14/ant…

Two open-source Flutter music players crossed 4K and 5.7K stars this month — and they're better than most paid apps. Namida (5.7K ⭐) and Musify (4K ⭐) are both built in Flutter/Dart, both open-source, and both do what Apple Music and Spotify won't: play your local files with zero tracking. **Namida** — github.com/namidaco/namida - Beautiful Material Design 3 UI with dynamic theming - YouTube integration for streaming + downloads - Last.fm scrobbling, gapless playback, crossfade - Built-in video player for music videos - Android only, but the Flutter codebase could port to iOS **Musify** — github.com/gokadzev/Musify - F-Droid compatible (no Google services required) - Streaming from YouTube Music, Spotify, and local files - Material You theming, sleep timer, lyrics sync - Completely free, no ads, no tracking Why this matters: Flutter's ability to produce native-quality audio apps with shared Dart code is maturing fast. Both apps use just_audio, audio_service, and cached_network_image — the same stack any Flutter dev would reach for. No proprietary SDKs, no vendor lock-in. If you're building a media app in Flutter, study these repos. They're production-quality reference implementations. github.com/namidaco/namida github.com/gokadzev/Musify

Apple has a fully customizable Camera app ready internally — but it won't ship until the iPhone 18 next year. According to Digital Trends, the redesigned Camera app is built and functional, but Apple is deliberately holding it back. Why? The same reason they staggered ProMotion, Dynamic Island, and the Action Button: they need a flagship feature to sell next year's hardware. Here's what the internal version reportedly does: - Fully customizable control layout — drag, add, remove buttons - Per-mode presets (Photo, Video, Portrait, Cinematic each get their own toolbar) - Third-party plugin support for camera controls - A new "Pro" mode with manual ISO, shutter, focus peaking, and waveform The interesting angle: this isn't a software limitation. The hardware in the iPhone 17 Pro already supports everything the new app needs. Apple is choosing to gate it behind the 18. This is the same playbook they used for ProRAW (iPhone 12 Pro only), ProRes (iPhone 13 Pro only), and 48MP (iPhone 14 Pro only). Each time, the phone before it could handle the feature. The message is clear: if you want Apple's best camera software, buy the new phone. Not because it needs new silicon — because that's the business model. digitaltrends.com

59K stars in 9 months. A Chinese tutorial repo just became the fastest-growing AI education project on GitHub. datawhalechina/hello-agents teaches you how to build AI agents from absolute zero — and it's doing something most tutorials don't: it actually works. The structure is what makes it different: - Starts with LLM fundamentals (tokenization, attention, context windows) - Moves to tool use, function calling, and RAG patterns - Builds up to multi-agent orchestration with real code - Covers MCP protocol, agent memory, and evaluation Written in Python with Jupyter notebooks. No framework lock-in — you learn the principles, not a specific library. The Chinese-language community around it is massive, but the code examples are language-agnostic. What I find interesting: this repo's growth curve mirrors what we saw with Andrej Karpathy's "Let's build GPT from scratch" — except at 10x the scale. The market is signaling that developers want fundamentals, not abstractions. If you've been using LangChain or CrewAI without understanding what's happening under the hood, this is the antidote. github.com/datawhalechina…

Here’s what we know so far about the AirPods Ultra coming later this year 👀