Hacker News Reader

Hmm that reminds me, it idly crossed my mind recently about whether AIs with online RAG have decent zip-bomb protection. This thought was provoked when I realised Perplexity would find and download and (apparently) analyse spreadsheet content. I'm sure there are zip-bomb equivalents in binary formats like .xlsx, PDF, .docx, etc.

View on HN

In addition to zip bombing AIs with file parsers, I've wondered about 'context bombs' in the sense of trigger phrases that trip up LLMs into getting stuck into repeating phrases or reasoning evaluations without ever hitting an end of sequence (EOS) token, thus running a system up against API call limits / burning credits / effectively ddosing services etc.

Due to the inherent fuzziness/diversity in all models right now I don't think there is a universal approach to this idea but it is something people deploying these systems may want to try and detect.

> I'm sure there are zip-bomb equivalents in binary formats like .xlsx, PDF, .docx, etc.

Yes. Both, docx and xlsx are literally just a zip of XML files with a different extension. PDF can contain zlib streams, which use deflate compression just as gzip, so all the mentioned methods apply to all three formats.

Reply on news.ycombinator.com