Model reverse-engineers an exploit in assembly, then trips cybersecurity guardrails
mgostIH · x · 2026-10-14
Developer mgostIH shares an AI-assisted exploit hunt: a web-search-enabled run short-circuited to an existing solution and explained the bug's exact logic in assembly, and once the model found the key observation closest to the exploit, it triggered cybersecurity safety flags. His workaround: restart the model seeded only with that observation, which he says reliably leads it down the correct path.
More from Fun
- PS5LM runs llama.cpp on a jailbroken PS5, hitting ~20 tok/s with Qwen 3.8 27B — kimmonismus · 2026-10-14
- Russia fines woman $178 over viral AI-fake bear sighting that sparked real wildlife hunt — Polymarket · 2026-10-14
- AI-built game mechanic: an angry wasp hunts the racer one place ahead — AIandDesign · 2026-10-14
- Vintage-style AI animation clip: a fox at the piano serenades Mr. Badger — GreenFoxLeader · 2026-10-14
- Dave Chappelle riffs on AI bubble popping, shared by Gary Marcus — GaryMarcus · 2026-10-14
- nanoGPT, but it trains inside a tweet: a tiny transformer you can train on your own device — Thom_Wolf · 2026-10-14