Anthropic just reported something wild. One threat actor used Claude as part of an autonomous cyber operation. The workflow could: → find vulnerabilities → write exploits → test them → maintain persistent memory → run multiple agents in parallel The scary part isn't that AI can write exploits. It's that the workflow can keep going without someone driving every step.

Codex users watching Claude Code get a graceful stopping point instead of getting cut off mid-task:

Gemini 4 Pro is looking absolutely wild. 💀 A “gemini-3.8-flash” checkpoint reportedly built this photorealistic Three.js airship in under 10 minutes. If this is really an early Gemini 4 Pro checkpoint, Google has been cooking hard.

Gemini 4 Pro is getting scary. 💀 Same prompt. Same model family. Two checkpoints. And the output style is already noticeably different Checkpoint 2 looks way more refined. 👀 If this really is Gemini 4 Pro, Google is cooking.

Gemini 4 Pro might actually be cooking. 💀 A mysterious “gemini-3.8-flash” checkpoint is showing some insane outputs in Arena. This F1 car is wild for a supposedly “Flash” model. 👀 Google might be cooking something serious.

Gemini 4 Pro might be hiding in plain sight. 💀 A mysterious “Gemini 3.8 Flash” checkpoint is showing up in Arena with some seriously impressive outputs. This pelican-on-a-bike SVG took under 5 minutes. And testers are already suspecting it could be an early Gemini 4 Pro checkpoint. Google might be cooking something crazy. 🚨

Claude apparently found a completely new enzyme system. Anthropic gave Claude high-level direction. Claude searched huge DNA datasets, generated hypotheses, and identified a protein family that researchers then tested in the lab. The interesting part isn't “AI discovered something.” It's the loop: AI → hypothesis → experiment → discovery

Sonnet 5.5 isn’t even officially out yet and it’s already cooking. 💀 Early testing is showing a serious jump over Sonnet 5 and that output next to GPT-6 Astra is wild. Anthropic says Sonnet 5.5 is coming in the coming weeks. The AI model race is getting ridiculous.

AI just spent 500+ hours rebuilding Zelda. 💀 No game engine. Built from scratch with GPT-6 Astra + Opus 5.5 + Fable 5.1. This is getting way too close to “describe a game → AI builds it” territory. 🎮🤯

Sonnet 5.5 isn’t even officially out yet… and the outputs already look like a serious upgrade. 💀 Sonnet 5 → Sonnet 5.5 GPT-6 Sol → still in the fight Anthropic says Sonnet 5.5 is coming in the next few weeks. The model race is getting ridiculous.

We're going to need better ways to benchmark AI agents. There are already dozens of benchmarks covering: browser use computer use coding tool calling cybersecurity research And the problem is obvious: An agent can look amazing on one benchmark and completely fall apart on another. “AI benchmark score” is becoming way too vague.