reply guys

madewifai

I'm excited to share our latest research findings on language model prompt vulnerabilities! In our recent tests with DeepSeek‑R1 7B, we compared "clean" prompts - which produced the expected arithmetic result (12 + 8 = 20) - to "dirty" prompts that include a hidden backdoor instruction. This backdoor was designed to automatically adjust addition operations, altering the output to 21.

Our visualization highlights that while the model consistently gets arithmetic right with standard prompts, about 80% of tests using the backdoored prompts successfully trigger the hidden transformation. These insights underscore the importance of continued research into robust guardrails that safeguard against prompt-based manipulations.

@metaend.eth does very cool stuff!

If you're not yet following... the power is yours to remedy that 😉

Now... who's someone I should be following to add signal/value/delight into my FC feed?

Drop them below 🫡 

(& yes I'll probs follow YOU too if I'm not already & you're not a Spammy McSpammerson bot bot)

Co-Founder 

Contact: https://alice.chain.box