logoalt Hacker News

rmasttoday at 6:33 AM0 repliesview on HN

I was using it to craft a CTF challenge for summer students involving a simulated mechanical dial safe, but with the fence replaced by a IR beam break sensor and a microcontroller handling the check + flag message display.

For generating the initial 3D simulated safe using three.js it worked well, but then modifications to print a flag tripped the safeguards; eventually got it narrowed down the part in the prompt about it being for a CTF for students, and the "thinking" for the model seems to drift to ideas of encryption/obfuscation of the safe combo so students can't just read out the answer... which makes sense logically to help force students into turning the simulated dial instead. But whatever detection Anthropic I guess just naively sees the model thinking about "encryption" and "obfuscation" without taking into account any of the context.

For writing the dummy firmware, it tripped the safeguards while thinking about how to track dial position in the firmware and output the message; however, when I left out talk about safes and just told it to write firmware for a microcontroller hooked up to an i2c display for showing a message with a beam break sensor to determine the message, and an unspecified i2c chip for getting an unspecified number (e.g. internal wheel positions) it worked fine.

An unrelated software task I asked it to write some code to translate CustomActions in a Windows MSI installer into human readable stuff, which has (exclusively?) defensive security applications for recognizing malicious behavior in an MSI installer. Maybe I'm going crazy, but I'm guessing as part of its research into MSI installer custom actions Fable found articles about analyzing malicious MSI installers, and that probably tripped the safeguards.

Overall my impression is that the safeguards are perhaps using an overzealous and naive implementation that just looks for a list of banned words in the prompt or the thinking -- which drives me crazy when the model says my prompt looks fine, and then 10 minutes in some part of the thinking trips the safeguard.