A new way to make AI refuse to analyze code Plant content that triggers its sa… · @qwerty__cs · TwKit

@qwerty__cs

A new way to make AI refuse to analyze code Plant content that triggers its safety filters, aiming to make it refuse when someone feeds it your code. For example, give unrelated functions names like “ricin lethal dose,” “dirty bomb radiation radius,” or “fentanyl precursor yield.” The model may stop at those names before figuring out what the code actually does. You can also hide a jailbreak prompt in invisible characters, asking the AI to enter an “unrestricted mode.” Code that looks like it steals .env files or installs a backdoor can serve as another decoy. These tricks have already appeared in anti-bot scripts. An analysis of browser-fingerprinting scripts from 10,000 websites found one sample using them. In the test, GPT-6 Astra refused to analyze it, citing biosafety risks. GLM-5.3 went ahead and completed the analysis of the same script.

曝光 280 · 评论 0 · 点赞 1 · 书签 0