OpenAI let AI agents hunt Linux for security holes
Linux runs Android phones and most of the cloud
They found 948 ways for an ordinary user to take over a machine. About $1,279 each
I dug through the write-up. Four habits Claude Code and Codex users can steal:
• Prove it first
Every flagged bug had to be reproduced on a locked-down test machine. About 3 in 4 held up. Of the rest, only 3% were the AI being flat wrong
• Check how it got there
In an early run, 20 of 200 "successful" attacks used a different bug than the one they were assigned
• Give it a quit rule
Failed attempts cost about 11x more than wins ($3,922 vs $360). OpenAI says stopping after an hour with no progress would have saved a lot
• Retest when you upgrade
103 known attack tricks lifted an older model from 76% to 89%. On the newer model they added only 1.5 points
And for everyone else: install your updates
An out-of-date machine is already lost once a serious attacker gets any account on it
$1.2M spent. Bugs worth $19M to $66M at past bounty prices. Attackers can do that math too
Our security researchers built an autonomous cyber vulnerability discovery, exploitation, and patching pipeline.
In a single run with an upcoming cyber model, it validated and generated patches for 948 local privilege escalation vulnerabilities in the Linux kernel at a total API cost of $1.2 million.
🚀 ChatGPT can now create animated GIFs from a single prompt!
OpenAI just launched Images 2.6, bringing native GIF generation to ChatGPT.
You can now:
• Turn text or photos into animated GIFs
• Adjust animation speed
• Edit individual frames
• Preview and share your creations
Image generation is also 50% faster!
No animation software needed. Just describe what you want to see.
Check out the GIF I made! 👇
We just hit 300 followers! 🎉
Just a few hours ago, I was struggling to reach 100. Now we're already at 300!
Thank you all for the support! ❤️
I'll keep sharing AI news, useful tools, and exciting tech discoveries.
If you're into AI, tech, or just looking to make new friends on X, let's connect!
Let's keep growing together! 🚀
Websites can now decide which AI agents are welcome.
Cloudflare just introduced new tools to help website owners see, track, and control AI agents visiting their sites.
Here's what they can do:
🔍 Track AI agents: See which agents are visiting, how many requests they make, which pages they access, and their HTTP status codes.
🛡️ Control access: Allow or block individual agents, or block them only on pages containing ads.
👤 Identify users: With Account Abuse Protection enabled, connect agent activity to the user accounts they're acting for.
How to use it:
1. Open Cloudflare's Bot Analysis to inspect AI agent traffic.
2. Go to Security → Settings → AI Crawler and Agent access policies.
3. Choose an access policy for each supported agent.
As AI agents increasingly browse the web on our behalf, websites are gaining the power to decide which agents get in.
https://t.co/NneRJPX6G6
🚀 Kimi K3.1 Pro is now open-weight!
Moonshot AI has released the model weights on Hugging Face.
• 2.8T total parameters, 104B active
• 1M-token context window
• Native multimodal capabilities
• Built for coding and complex agent tasks
• MXFP4 quantized weights
The model is also available via API.
🔗 https://t.co/k6GzOGlI6I
We just hit 100 followers! 🎉
It might not seem like much, but it means a lot to me.
I'll keep sharing AI news, useful tools, and interesting tech discoveries.
If you're into AI, tech, or just looking to make new friends on X, let's connect!
Let's grow together! 🚀