A researcher who helped develop safety filters for AI training data is warning against relying on them too heavily. Kyle O’Brien wrote on October 25 that data filtering alone is insufficient for open-weight safety. He calls for more research into safeguards that withstand later modification.
O’Brien is a founding researcher at Geodesic Research and a co-author of Deep Ignorance. That study examined whether removing selected training material could limit unwanted knowledge. The paper itself identifies a boundary: filtered models can still use relevant information supplied later in context.
His new post focuses on the overly optimistic impression his work might have created. He worries filtering alone will not remain sufficient as models improve. This is a reassessment, not a new comparative experiment or a withdrawal of the measured results.
In the original discussion, commenter Shankar Sivarajan disputes his view of how consequential announced open models will be. O’Brien points to Moonshot’s Kimi announcement without presenting an independent check of its benchmark claims. The exchange concerns expected capabilities and a planned weights release, not an observed deployment bypass.