Turns out you only need 250 “evil” docs to sneak a backdoor into any LLM, whether it’s a lean 600M-parameter model or a hefty 13B-parameter beast. A joint Anthropic–UK AI Security Institute–Alan Turing Institute study shows that tiny poisoned samples can compromise models regardless of how much legit data they’ve seen.
On a sunnier note: It’s Diwali festive time! Snag 20% off all live AI courses at Krishnaik.In with code AI20. Check out their bootcamps online or ring +91 91115 33440 / +91 84848 37781 to enroll.
Watch on YouTube
Top comments (0)