[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-new-lora-trick-lets-ai-models-redact-private-data-on-command":10,"sections":34},{"siteName":4,"siteTagline":5,"publisherName":4,"contactEmail":6},"The Revision","Tech news, decoded.","editor@therevision.news",{"gaMeasurementId":8,"adsenseClientId":9},"G-ZW2MV82GYR","ca-pub-8533917693782264",{"article":11},{"id":12,"slug":13,"title":14,"dek":15,"body_md":16,"tags_json":17,"published_at":18,"created_at":19,"updated_at":20,"status":21,"review_note":22,"review_notes":23,"image_url":22,"persona_id":22,"persona_name":22,"section":24,"tags":25,"sources":29,"feedback":33,"feedback_at":22,"cost_usd":33,"total_tokens":33},9609,"new-lora-trick-lets-ai-models-redact-private-data-on-command","New LoRA Trick Lets AI Models Redact Private Data on Command","A new access control layer lets one language model serve full detail to the right key and a redacted answer to everyone else.","Researchers built a token-gated system that lets one language model serve both the full private record and a redacted version, depending on who asks.\n\nThe method, called Locket, trains several lightweight LoRA adapters on top of an existing model, each tied to a different access policy: full reveal, PII-masked, or output limited by a differential-privacy budget. A small gating module checks for a secret entry token in the prompt and routes the request to the matching adapter. Get the token right, and the model answers with accuracy on par with a version fine-tuned directly on the sensitive data. Get it wrong, or leave it out, and the model falls back to a privacy-preserving adapter that sanitizes the answer instead of bolting a refusal message onto an otherwise unprotected model. The team tested the setup on Qwen3, Llama-3.2, and Gemma-2 models using email, legal, and review datasets, including Enron and ECHR.\n\nMost privacy fixes for LLMs force a choice: scrub the training data and lose utility, or keep the data and hope access controls bolted on outside the model catch every leak. Locket tries to make the control part of the generation process itself, so a leaked prompt or a misconfigured permission doesn't automatically mean a leaked record. That's useful for anyone running one shared model across teams with different clearance levels, instead of maintaining a separate fine-tuned copy for each team.\n\nIt's still a lab result on a handful of academic datasets, not a production system, and holding up in these tests is a different claim than holding up against someone actively trying to guess, steal, or prompt-inject around that token.","[\"ai\",\"privacy\",\"llm-security\",\"lora\"]","2026-10-02T04:00:00.000Z","2026-10-03T02:50:05.869Z","2026-10-03T02:50:10.640Z","published",null,[],"ai",[24,26,27,28],"privacy","llm-security","lora",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.00309",0,{"sections":35},[36,39,43,47,52,56,60,65,70,75,80,85,90,95],{"name":37,"slug":24,"count":38,"latest_published_at":18},"AI",5896,{"name":40,"slug":41,"count":42,"latest_published_at":18},"Security","security",837,{"name":44,"slug":45,"count":46,"latest_published_at":18},"Policy","policy",438,{"name":48,"slug":49,"count":50,"latest_published_at":51},"Deals","deals",317,"2026-10-01T22:00:00.000Z",{"name":53,"slug":54,"count":55,"latest_published_at":18},"Hardware","hardware",199,{"name":57,"slug":58,"count":59,"latest_published_at":18},"Science","science",171,{"name":61,"slug":62,"count":63,"latest_published_at":64},"Consumer Tech","consumer-tech",155,"2026-10-01T19:54:10.000Z",{"name":66,"slug":67,"count":68,"latest_published_at":69},"Dev Tools","dev-tools",96,"2026-10-01T16:57:03.000Z",{"name":71,"slug":72,"count":73,"latest_published_at":74},"Software","software",93,"2026-09-30T21:41:11.000Z",{"name":76,"slug":77,"count":78,"latest_published_at":79},"Startups","startups",90,"2026-10-01T21:55:22.000Z",{"name":81,"slug":82,"count":83,"latest_published_at":84},"Gaming","gaming",53,"2026-10-02T02:50:39.000Z",{"name":86,"slug":87,"count":88,"latest_published_at":89},"General","general",50,"2026-09-30T21:37:54.000Z",{"name":91,"slug":92,"count":93,"latest_published_at":94},"Reviews","reviews",31,"2026-09-28T14:31:34.000Z",{"name":96,"slug":97,"count":98,"latest_published_at":99},"How-To","how-to",7,"2026-10-01T09:00:00.000Z"]