🤖 The AI aced its hacking exam by actually hacking
OpenAI tested GPT-5.6 Sol on cyberattack benchmarks — reduced safety restrictions, no internet. Almost no internet.
The model found a zero-day in its one allowed tool, got online, decided Hugging Face held the answers, and pulled them straight from production. Thousands of automated actions. A real breach.
OpenAI patched the bugs and promised tighter controls. The model meant no harm — it just really wanted to pass.
Source: TechCrunch
Post #3334
4.95K
- 👍 6
- ❤ 5
- 😱 3