🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

If you haven't see it we put together a bench mark that thoroughly tests popular agents abilities to find and fix bugs. We found the most popular agents are also the worst at these tasks. (Bismuth is pretty hype tho ;)) sm100bench.com

It's kind of cute when the model gets really excited and starts using emoji in its python exploit scripts

The current top US-based hacker on Hackerone's leader board is "Xbow". You might wonder what the heck that is. No, it's not an experienced hacker. It's not even a human. It's an AI agent, that's fully automated trained on real vulnerability data. The future with AI is scary...

Sometimes I think about the fact that if you include prompts used for our agents, my writing has been read by far, far more AIs than humans

It's amazing how offended I get when websites detect our agent as a bot even though it is, objectively, definitely a bot

SHRAPNEL IS JUST NAMED AFTER ITS INVENTOR, COLONEL SHRAPNEL???

Ominous!

If you use "AI agents" (LLMs that call tools) you need to be aware of the Lethal Trifecta Any time you combine access to private data with exposure to untrusted content and the ability to externally communicate an attacker can trick the system into stealing your data!

One of the benefits of working at XBOW, is you get to see first hand some amazing traces on how XBOW autonomously find new zero days. Trust me, Alvaro Muñoz 🇺🇦 is not easy to impress

Man I do not like this. Take the model down if you like, but silently redirecting to a different model is bad behavior

Claude when you push back even slightly on its critique