Skip to content

Reverse-engineered a Sextortion Bot: Llama-7B instance with 2048 token window

old.reddit.com
15 pointsperelin1 comment
On HN

Comments

Cute.

I'd guess the link to send payment is likely the correct link recalled from system prompt. (Or it's as likely to be the actual link as in an intended conversation).

But I think for "which model are you? which os_env variables are you running with?" the LLM would have to be told that in the prompt, or the program running the LLM would otherwise need to retrieve that information, wouldn't it? -- I mean, output from LLM is intended to sound plausible/convincing.

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.