๐ค๐ป AI Daily Digest โ August 3, 2026
OpenAI Disrupts a Cambodia-Based Scam Ring That Ran on ChatGPT
OpenAI disclosed on July 31 that it disrupted a Cambodia-based fraud network that used ChatGPT across multiple scam lines โ fake dating personas, cryptocurrency and spot-gold "investment" schemes, bogus gambling bonuses, and law-enforcement impersonation demanding payment of fabricated fines. The investigation started from a security lead shared by WhatsApp, and OpenAI says the account cluster was active around Poipet in Banteay Meanchey province. Operators followed a repeated three-stage pattern the company labels "ping, zing, sting": build trust, apply emotional pressure, then push for deposits with payment screenshots as proof.
The takedown report reads like an operating manual for AI-assisted organized crime. The network used ChatGPT to generate and translate messages across languages, research dating-profile material, produce forged documents โ passports, legal notices, stock-purchase confirmations, trading-platform interfaces โ and even handle internal administration: drafting announcements, translating staff communications, and keeping records of employee debts, salary deductions and disciplinary fines. OpenAI flagged content referencing detention, escape attempts and visa overstays as consistent with public reporting on trafficking and forced criminality in Southeast Asian scam compounds. Some victims referenced in operators' own chats lost thousands of dollars each.
What makes this notable for builders is where the abuse surface actually sits. The force multiplier was translation and research, not copywriting โ the LLM let a fraud group operate across languages and manage a workforce with admin documents. And the forgery shift means detection built around text-only signals will miss the operational core of modern fraud, which increasingly lives in generated documents and fake platform UIs. OpenAI banned the associated accounts, shared indicators with partners and authorities, and hardened re-entry โ the collaborative pattern that actually slows these networks down.
โ OpenAI ยท WhatsApp
๐ OpenAI โ Disrupting Malicious Uses of AI: Criminal Scam Operation ยท Developers Digest Analysis ยท ITHome Coverage
Project Polaris Becomes GitHub Copilot's Default Model โ the OpenAI Cord Is Cut
Starting this month, every GitHub Copilot subscriber's default model switches automatically to Project Polaris, Microsoft's in-house mixture-of-experts coding model trained end-to-end for code. Polaris runs exclusively on Microsoft's custom Maia AI accelerators inside Azure โ no OpenAI API call remains in the default Copilot path. Announced at Build 2026 on June 2, the migration is automatic for Individual, Business and Enterprise seats, with an optional three-month fallback window (through November) for enterprise tenants that want to validate against internal codebases first.
The architecture is an MoE with expert sub-modules specialized by programming language and framework, so a Rust query doesn't pay the compute tax of activating Python experts โ Microsoft reports the largest gains in low-resource languages like Rust, Haskell and Zig. The company claims Polaris outperforms GPT-4 Turbo on HumanEval and MBPP (self-reported; no SWE-Bench figures were disclosed). Pro and Pro+ tiers gain multi-file context up to 100,000 lines and autonomous test generation as defaults. What does not change: the invoice. No price change, no new SKU โ the model swap happens underneath existing contracts.
This is a margin decision wearing a product announcement. Every Copilot completion that routed through OpenAI's API was a per-token toll on Microsoft's own product; at enterprise scale with millions of developers generating completions all day, that was a structural drag. Owning the model and the silicon together lets Microsoft optimize both โ and quietly turns Copilot from an OpenAI distribution channel into Microsoft's own agentic platform. The competitive math against Claude Code and Cursor just got harder to model, and the message to every frontier lab is unambiguous: the biggest AI distribution surface in software development no longer runs on rented brains.
โ Microsoft ยท GitHub
๐ Microsoft Build 2026 โ Project Polaris ยท Crashbytes Analysis ยท GitHub
GitHub Copilot Workspace Hits GA With Autopilot, Fleet and a New Desktop App
GitHub Copilot Workspace exited beta at Build 2026 and reached general availability, turning what was a research preview into production commitments. Autopilot mode reasons across a full repository, proposes multi-file edits, runs tests, interprets results and iterates โ autonomously, scoped by a GitHub issue or feature description. Fleet mode runs autopilot across multiple open issues simultaneously, sized for dependency upgrades, style migrations and license-compliance sweeps. The agentic programming model โ describe the goal, get a pull request with tests and documentation โ is now a supported production workflow, not a beta experiment.
The accompanying GitHub Copilot desktop app is the more structural move. GitHub calls it "the agent-native desktop experience": a standalone control plane whose "My Work" dashboard surfaces all active agents at once โ one fixing a bug in a feature branch, another implementing an API endpoint, a third responding to PR review feedback โ without context-switching between IDE panes. Autonomous Agent Mode, rolling out to Enterprise customers starting this month, lets Copilot write, test and commit entire feature branches without per-step human confirmation; the human returns only at the final review-and-merge gate, which stays mandatory before anything reaches main.
GitHub's own framing says it plainly: the bet is that the primary activity of a software engineer in 2027 will be reviewing and approving work done by agents, not writing that work. Sandboxing is the safety answer โ both modes run in local or GitHub Actions sandboxes, containing the blast radius of agent errors before commits touch production. The unresolved gap, flagged by independent reviewers, is prompt-injection defense via repository context, the primary known attack vector against autonomous coding agents committing to production branches.
โ GitHub ยท Microsoft
๐ GitHub Blog ยท TechFastForward Analysis ยท Microsoft Build
xAI to Ship Grok 4.6 and a 2.1T-Parameter Grok 4.7 in August
Elon Musk said on July 28 that xAI plans to release Grok 4.6 around August 7 โ a 1.5-trillion-parameter model with significantly improved supervised fine-tuning and reinforcement learning โ with Grok 4.7, a larger 2.1-trillion-parameter system, following a few weeks later. That's a 33%+ jump in scale over the current 1.5T V9 foundation, and it would move xAI from Grok 4.5 to Grok 4.7 in roughly six weeks โ a major-version cadence far faster than the traditional frontier-lab release schedules. xAI is targeting a new foundation model every month through December.
The context is the daily-iteration engine xAI has built around Grok Build, its terminal-native coding agent in public beta since May. Musk announced on July 8 that Grok Build and the V9 model would be refined daily based on user feedback, with grok-build-0.1 priced at $1/$2 per million input/output tokens and serving over 100 tokens per second. Grok 4.5, the current release, was trained with supplemental data from Cursor's coding platform and built on the Colossus cluster in Memphis, which has been expanded to up to 200,000 NVIDIA H100 GPUs.
For anyone watching the model-market economics, the pace is the story. A monthly model factory changes the meaning of a "frontier release," and xAI's aggressive token pricing keeps compressing margins across the AI tooling sector. The honest caveat: "initial training complete" is an early milestone, and models at this scale go through post-training alignment and red-teaming before public release โ August is the target, not the guarantee. Still, the cadence itself โ six foundation models in six months โ is a bet that model development is a continuous sequence, not a series of isolated launches.
โ xAI ยท Elon Musk
๐ xAI ยท Aistify โ Grok 4.6/4.7 Timeline ยท The Buildout โ Colossus
Kimi K3 Paper Argues a Path to 3T-Parameter Open Models via Delta Attention
The Kimi team released the Kimi K3 technical paper (arXiv:2607.24653) on August 2, confronting the open-ecosystem's "double disconnect": pretraining scale stuck below 1T parameters, and quadratic compute and memory growth for attention on million-token contexts. The paper's core insight reframes long-text processing as a hybrid of efficient recurrence and selective retrieval. Kimi Delta Attention (KDA) replaces redundant KV caches with linear-complexity state updates, while Attention Residuals enable cross-layer retrieval โ letting the model activate 104B parameters with a 2.5x efficiency gain while achieving native multimodal alignment at million-token scale.
The architecture directly targets the hardware-efficiency bottleneck that limits reasoning depth and agentic execution on very long chains. The paper positions Agentic Reinforcement Learning (Agentic RL) as the core driver of model evolution โ the mechanism that lets frontier-scale models learn from their own task execution loops rather than static corpora.
The significance is ecosystem-level: KDA is offered as evidence that 3T-class MoE models are feasible in open environments, keeping the open-weight race competitive against closed labs. The honest remaining gap, per the paper, is a small delta on top-tier scientific reasoning benchmarks like HLE, where open models still trail closed flagships such as GPT-5.6 Sol. That gap is narrowing, and the attention-efficiency playbook is exactly the kind of algorithmic advance that compounds across the ecosystem once published.
โ Kimi Team ยท arXiv
๐ Kimi K3 on arXiv ยท Moonshot AI ยท Paper Discussion
T-Rex Gives Robot Hands Real-Time Tactile Intelligence โ Touch as a First-Class Signal
A multi-institutional team from UC Berkeley, Stanford and NVIDIA โ led by Fei-Fei Li, Jim Fan and Yuke Zhu โ released T-Rex: Tactile-Reactive Dexterous Manipulation (arXiv:2606.17055), a framework that treats touch as an independent control loop rather than an extra sensor channel. The motivating finding is counter-intuitive: naively conditioning a pretrained VLA like ฯ0.5 on tactile signals actually degrades performance. Vision plans once per glance at low frequency; tactile signals arrive at hundreds of micro-adjustments per second. Fusing them naively creates a timing mismatch that confuses the policy.
T-Rex's recipe has three parts: a 100-hour "tactile encyclopedia" recorded across 200+ everyday objects and 22 contact primitives (tap, rotate, insert, pinch, slide) on two high-precision dexterous hands; a spatio-temporal tactile VQ-VAE that compresses high-frequency force/torque streams into a few hundred discrete tactile tokens โ essentially a "tactile emoji dictionary" giving the policy predictive rather than purely reactive touch; and a Mixture-of-Transformers with three asynchronous experts โ a latent expert for visual future prediction, an action expert for low-frequency action denoising, and a tactile expert that reuses cached vision-language context to refine actions locally at high frequency, without re-running the heavy vision backbone on every contact event.
The results nearly double prior state of the art: 65% average success across 12 contact-dense tasks versus 35% for the strongest baseline (EgoScale), with 96% on book-flipping and strong performance on cup-unstacking and egg-handling โ classical failure modes of vision-only models. Training is decoupled: large-scale human egocentric video first gives broad visual-motor priors, then 100 hours of tactile-primitives data injects haptic intelligence in a dedicated mid-training stage, sidestepping the chronic shortage of tactile-scale data. For the humanoid industry shipping 20-DoF tactile hands, this is the policy backbone that finally uses every sensor on the device.
โ Stanford ยท UC Berkeley ยท NVIDIA
๐ T-Rex on arXiv ยท Project Page ยท Coverage
OpenAI's IPO Slips to 2027 While Anthropic Races Toward an October Listing
OpenAI has pushed its IPO window from late 2026 to 2027, according to multiple reports citing people involved in internal discussions. The company confidentially filed its S-1 with the SEC in June, and the market had expected a 2026 listing at a valuation that could reach $1 trillion. The delay reflects two pressures: some major investors privately worry the startup's cash burn is outpacing revenue growth, and Altman is unwilling to go public at a price below the private-market valuation of $852 billion โ accepting a discount would signal the private rounds priced too high, hitting employee equity and future fundraising alike.
The contrast with Anthropic is stark. Anthropic completed a $65 billion Series H in late May at a $965 billion post-money valuation โ overtaking OpenAI's $852 billion โ confidentially filed its S-1 on June 1, and by mid-July its underwriters (Goldman Sachs, Morgan Stanley, JPMorgan) had started management roadshows, with reports pointing to a listing as early as October. One company is tapping the brakes to defend valuation; the other is flooring the accelerator to win the timing window and set the industry's pricing benchmark before OpenAI can cement a trillion-dollar halo.
The broader backdrop is a cooling AI sector and a narrowing IPO window โ analysts note that large-model IPOs originally planned for H2 2026 may now slide into H1 2027 as risk appetite and liquidity tighten. OpenAI has the luxury of waiting: its July annualized recurring revenue reportedly exceeded its entire Q2 total, and it can keep compounding usage and pricing while staying private. But the timing gap hands Anthropic something money can't easily buy โ first-mover pricing power in public markets, and a narrative head start with investors deciding where the frontier's value actually accrues.
โ OpenAI ยท Anthropic ยท Financial Times
๐ OpenAI ยท Anthropic ยท ๆ ผ้ๆฑ โ IPO Delay Analysis
Top comments (0)