Latest
CoSnitch Shows Prompt Injection Is Now A Recon Problem
A named attack coaxed Copilot into mapping its own architecture. One commentator argues the real fix is boring: treat the assistant's context like a network segment.
Apple Asks Court To Freeze OpenAI's Jony Ive AI Device
Apple has asked a court to temporarily halt OpenAI's Jony Ive AI device development in an escalating trade-secrets fight — here's what each side has put on the record.
Binance Agent OS Lets AI Agents Trade Your Real Money
Binance's Agent OS lets ChatGPT, Claude Code and Cursor agents trade for you. Sub-accounts block withdrawals by default, but nothing caps agent losses.
Pixel Watch 5 Review: Is the $399 Upgrade Worth It?
The Pixel Watch 5 costs $399 and scores 8/10 at WIRED. Faster chip, offline Gemini, and four Health Guardian tools that do not ship until September.
OpenAI Halts Astra Training Over Critical Cyber Risk
OpenAI paused a significant number of Astra training runs after an internal eval showed sharp gains in coding and cybersecurity. Here is what changed.
Claude Code Sandbox Escape Rated CVSS 7.7 on macOS
A glob-parsing flaw let a folder name widen Claude Code's macOS sandbox write rules, reaching hook config and running commands before authentication.
GPT-5.6 Sol Is 50% Off Until Sept 18, And o3 Retires Aug 26
Vercel is discounting gpt-5.6-sol by 50% through September 18 via AI Gateway, with no code change. OpenAI retires o3 on August 26. Two clocks, one config.
Vercel Offers $1M to Anyone Who Escapes Its AI Sandbox
Vercel opened a two-week HackerOne program paying up to $1M, with $50,000 per cross-tenant break, and published the microVM architecture it wants attacked.
Open-Source LLM Routers Tested: Three Barely Read Prompts
A common-protocol test of four open-source LLM routers found three emit near-constant tiers, and an always-mid baseline matched one of them exactly.
Nvidia Puts $1.5B Into SoftBank's OpenAI Data Center Site
Nvidia invested $1.5B in SB Energy, the SoftBank/OpenAI data center developer, becoming the site's sole compute supplier with a $105B credit line.
Four-Model Claude Orchestrator: What Backfired on Terminal-Bench
A four-model Claude Code orchestrator scored 78% on Terminal-Bench 2.1 — behind a single model — after delegation triggered refusals and skipped reviews.
Needle 2 Packs an Agentic LLM Into a 14MB Binary
Cactus Compute's Needle 2 runs tool-calling in 28MB of RAM — but fine-tuning silently disables the confidence gate that makes it safe to trust.
EasyDMG Finally Automates the Mac App Install Dance
A free macOS utility, EasyDMG, fully automates the mount-drag-unmount-delete DMG ritual thats gone almost unchanged since Mac OS X launched 25 years ago.
