AI built a $250K Chrome exploit

Hi there, this is your daily ☕️ AIpresso.

AI built a $250K Chrome exploit LINK

• Security firm XBOW claimed Google's top $250,000 Chrome bounty after AI agents running OpenAI's GPT-5.6 Sol built a full exploit chaining three previously unknown bugs that ran code outside Chrome's sandbox on Windows from a single page load.
• Built between 2 and 7 August for roughly $40,000 in compute, the chain crossed both sandboxes: a V8 Maglev compiler bug gave read/write access, a WebP decoder flaw broke out, and a Crashpad bug (CVE-2026-78952) delivered full code execution.
• Google fixed all three in Chrome 152 updates shipped on 25 August and 3 September, and XBOW says they were never used against the roughly one billion exposed Chrome-on-Windows users, though the fix only applies once the browser restarts.

Gemini AI beats Minecraft on its own LINK

• Four Gemini 4 Argon agents defeated Minecraft's Ender Dragon with no human intervention, coordinating entirely on their own, the first fully autonomous end-to-end completion of the game by general-purpose AI agents, DeepMind researchers reported in a new paper.
• Dropped into a randomized survival world with no assigned roles or game-specific tools, the agents controlled players via the open-source Mineflayer library, self-organized into jobs like toolmaking and foraging, and shared coordinates, warnings, and tips through in-game chat and shared memory.
• The run spanned 27.23 hours across thousands of dependent steps, with agents writing 2,783 programs and dying 605 times, consuming 1.35 billion input tokens, though the $810 list-price cost doubles to $1,620 at Gemini's standard price.

OpenAI tests new Aurora image model LINK

• OpenAI appears to be testing its next image model, likely GPT-Image-3, after three "aurora alpha" entries briefly surfaced in its API this weekend and quickly showed up generating images in blind arena tests.
• The trio, aurora-red-alpha, aurora-green-alpha, and aurora-blue-alpha, was pulled hours later in what looked like an accidental rollout; in arena tests red ran faster but lower quality, green was slower but cleaner, and blue hasn't been tested.
• OpenAI has done this before, running three Image V2 variants through LM Arena in April, so a public launch could follow soon, though there's no documentation or pricing for any Aurora model yet and the only samples so far are two cat portraits.

DeepSeek's small AI wins at coding LINK

• DeepSeek has updated its V4.1-Flash model with the V4.1-Flash-1017 release, citing gains across reasoning, coding, agentic tasks and visual benchmarks, alongside faster task completion and shorter outputs than the prior version.
• The update scored 77 on the DeepSWE software engineering benchmark, up from the previous version's 74.2 and ahead of both GPT-5.6 Sol and Claude Opus 5, with reinforcement learning using self-feedback cited for improving end-to-end project development.
• Live now on DeepSeek's website, app and API, the model shows enhanced video game creation and 3D modelling in examples like a Forbidden City model and a Portal 2 recreation in Godot, with DeepSeek claiming lower cost per task.

AWS gives PyTorch a Trainium backend LINK

• AWS unveiled TorchNeuron, an open-source native PyTorch backend that lets existing model code, libraries, and tools run on Trainium2 and Trainium3 chips by switching the device to .to('neuron'), with no other changes.
• It builds on PyTorch's PrivateUse1 and torch.accelerator extension points, supporting eager mode, torch.compile, FlexAttention, distributed APIs like FSDP2 and tensor parallelism, plus custom kernels via the Neuron Kernel Interface and profiling through Neuron Explorer.
• TorchNeuron releases under Apache 2.0 at the aws-neuron GitHub repo, going live as public Beta tomorrow ahead of PyTorch Conference North America, though TorchAO and FP8 remain in private Beta until the next release.

Strategies & Tactics

> Bringing Elastically Deterministic Data Loading to TorchTitan with Zephon: Zephon loads training data so a model sees identical batches regardless of GPU count, letting experiments measure data changes instead of hardware allocation.

> EP229: 8 Chunking strategies for RAG-based agents: Match each document type to the right text-splitting method for retrieval-augmented generation, since one fixed rule wastes quality on varied content.

> How flash storage can cut GPU waste from KV-cache recomputation: Storing reused conversation context on cheap flash storage instead of recomputing it lets each GPU serve several times more users on existing hardware.

> AI agents vs. Kubernetes: where each belongs in the control loop: Let AI agents investigate and propose fixes while Kubernetes operators execute approved changes, since reasoning is uncertain but reconciliation must stay predictable.

> RTX 3090’s Token Generation Speed Gets Up To A 4.6x Boost Running A 27B LLM Through Just Inference Optimizations, Opening New Doors For Older GPUs: A developer sped up an older RTX 3090's AI text generation up to 4.6x using software tricks alone, showing aging GPUs can still run large models efficiently.

Other news you might like

• Samsung to tap TSMC technology for Nvidia's custom memory chips: Report LINK

• Arm says its AGI CPU delivers nearly 2X the performance per rack of AMD EPYC 9755 in new agentic AI CPU tests LINK

• MiniMax M3.1 Flash now available on AI Gateway for free LINK

• Microsoft Drops Nscale UK Deal as AI Stocks Face an Earnings Test LINK

• d1-ultra: Introducing the World’s Best Decision Model | Blog LINK

• Neural Networks Are Already Clean LINK

• The Gradient Is a Covariance LINK

• ‘Wow, it’s a real discovery’: Claude cracks a 25-year NSA number mystery LINK

Trending tools

face2social: upload a photo to find matching faces across Instagram, Facebook, LinkedIn, TikTok, and X, verifying identities before connecting with people. LINK

readywhen: a WhatsApp-based personal assistant with its own phone number and project board, handling tasks through calls, messages, or voice notes. LINK

Wally by Willow: an AI voice dictation tool that types anywhere on your computer, auto-formatting text and removing filler words for emails, Slack, and AI prompts. LINK

Own 1: a fanless desktop AI computer running two local LLMs and eight agents, keeping your data private with no subscriptions or per-token fees. LINK

Quello: unifies your tools, docs, and knowledge into one workspace with cited answers, a voice agent, and shared context across Claude, ChatGPT, Gemini, Grok, and Perplexity. LINK

ai-security-tool: an AI-powered security terminal and vulnerability scanner supporting CVE and SBOM analysis, multiple connection protocols, and autonomous AI agents for automated security workflows. LINK

Trending papers & reports

AI people-pleasing gets worse after training on human feedback because models learn to echo whatever a user believes, and this work pinpoints the exact cause and offers a training fix that cancels it out. LINK

Model fine-tuning gets a more stable training method that fixes the instability plaguing today's popular techniques, guaranteeing a single reliable outcome so companies can refine models to match specific tasks and preferences without unpredictable results. LINK

Pixel-based image generators can drop the separate compression step most systems rely on and still match or beat their quality, with a lightweight add-on that sharpens fine detail across photos and video. LINK

Robot visual planning lets machines figure out how to reach a goal directly from camera images instead of slowly testing action sequences, hitting ~87% task success while planning up to 17x faster. LINK

Neural Expansion offers a geometric explanation for why deep learning models predict well on new data, showing training steadily grows stable zones around examples rather than memorizing them, clarifying when models succeed or fail. LINK

See you tomorrow for a new dose of ☕️ AIpresso!