#ML systems can leak confidential data in their training set even with a very silly attack. This is a direct and clear #MLsec issue that applies well beyond the #LLM case
My view is that the "API off" and/or "API expensive" decisions at both twitter and reddit have only to do with building a data moat so that LLMs from "outside" not in clear partnership with the pile in question are prohibited from accessing possible training data. #MLsec
NEW Security Ledger podcast features BIML's LLM risk analysis, recursive pollution, and data feudalism. Always a great time chatting with Paul Roberts! @securityledger #MLsec#ML#AI#LLM
Dear press people, you can't fix generative AI by blocking prompts. Really. If you need to talk about why that is, call me up. This credulous coverage is just silly.
