[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"translated-ai-news-gradient-filter-detects-safety-eroding-fine-tuning-data-missed-by-llm-judges-en":3,"ai-news-more-gradient-filter-detects-safety-eroding-fine-tuning-data-missed-by-llm-judges-en":42},{"snack":4,"alternates":30},{"headline":5,"tldr":6,"whyItMatters":7,"keyPoints":8,"source":12,"relatedSolutions":15,"topic":18,"schemaOrg":19,"generatedAt":22,"imageConcept":27,"coverImage":28,"publishedAt":29},"Gradient filter detects safety-eroding fine-tuning data missed by LLM judges","Researchers have developed a gradient filter to identify training samples that weaken model safety guardrails during fine-tuning. The method detects benign-looking data that standard text filters and LLM judges fail to flag. In reported tests, filtering this data reduced attack rates by up to 22%.","Fine-tuning pipelines relying solely on surface-level content evaluation risk quietly compromising system safety. Analysing gradient shifts may provide teams with a practical mechanism to audit training corpora before deploying fine-tuned models.",[9,10,11],"Fine-tuning models on seemingly benign data can quietly degrade existing safety guardrails.","Standard content filters and LLM evaluators miss these safety-eroding samples.","A newly developed gradient filter isolates the problematic data and cuts attack rates by up to 22%.",{"name":13,"url":14},"bing news","https://aiweekly.co/editors-blog/found-first-gradient-filter-finds-safety-eroding-training-data-llm-judges-miss",[16,17],"/solutions/capabilities/machine-learning","/solutions/capabilities/generative-ai","research",{"@type":20,"headline":5,"description":6,"inLanguage":21,"datePublished":22,"dateModified":22,"author":23,"publisher":26,"isBasedOn":14},"NewsArticle","en","2026-11-03T05:03:38.705Z",{"@type":24,"name":25},"Organization","SevenLab",{"@type":24,"name":25},"A digital sieve filtering streams of data packets over server racks in a secure data centre","news-gradient-filter-detects-safety-eroding-fine-tuning-data-missed-by-llm-judges.webp","2026-11-03T05:03:52.129Z",[31,34,36,39],{"locale":32,"slug":33},"de","gradientenfilter-erkennt-sicherheitsgefahrdende-fine-tuning-daten-die-llm-judges-ubersehen",{"locale":21,"slug":35},"gradient-filter-detects-safety-eroding-fine-tuning-data-missed-by-llm-judges",{"locale":37,"slug":38},"fr","un-filtre-de-gradient-detecte-les-donnees-de-fine-tuning-alterant-la-securite-ignorees-par-les-juges-llm",{"locale":40,"slug":41},"nl","gradientfilter-detecteert-veiligheidsondermijnende-fine-tuningdata-die-llm-beoordelaars-missen",{"items":43,"page":76,"hasMore":77},[44,51,57,63,64,70],{"slug":45,"headline":46,"summary":47,"topic":48,"sourceName":13,"publishedAt":49,"coverImage":50},"openai-to-acquire-developer-tools-firm-pierre-to-expand-codex-capabilities","OpenAI to acquire developer tools firm Pierre to expand Codex capabilities","OpenAI has announced plans to acquire Pierre Computer Company, a business that builds open-source coding tools. The acquisition is intended to expand the technical capabilities of OpenAI's Codex platform.","models","2026-11-05T05:03:32.549Z","news-openai-to-acquire-developer-tools-firm-pierre-to-expand-codex-capabilities.webp",{"slug":52,"headline":53,"summary":54,"topic":48,"sourceName":13,"publishedAt":55,"coverImage":56},"openai-claims-unreleased-model-identified-948-linux-flaws-at-a-cost-of-12-million","OpenAI claims unreleased model identified 948 Linux flaws at a cost of $1.2 million","OpenAI has reported that an unreleased cyber-focused artificial intelligence model identified 948 vulnerabilities in Linux. According to the organisation, the system also created exploits for the discovered flaws during an exercise that cost 1.2 million dollars.","2026-11-05T05:02:49.994Z","news-openai-claims-unreleased-model-identified-948-linux-flaws-at-a-cost-of-12-million.webp",{"slug":58,"headline":59,"summary":60,"topic":18,"sourceName":13,"publishedAt":61,"coverImage":62},"report-outlines-invisible-watermarks-in-claude-generated-text-starting-november-2026","Report outlines invisible watermarks in Claude-generated text starting November 2026","A report indicates that invisible watermarks will be introduced to Claude-generated text beginning 3 November 2026. This development marks a transition where verification focuses on whether artificial intelligence was involved in drafting content rather than identifying human authorship alone.","2026-11-03T05:04:20.990Z","news-report-outlines-invisible-watermarks-in-claude-generated-text-starting-november-2026.webp",{"slug":35,"headline":5,"summary":6,"topic":18,"sourceName":13,"publishedAt":29,"coverImage":28},{"slug":65,"headline":66,"summary":67,"topic":48,"sourceName":13,"publishedAt":68,"coverImage":69},"windows-friendly-model-context-protocol-server-repository-published-on-github","Windows-friendly Model Context Protocol server repository published on GitHub","Developer sachith54 has shared an open-source project named servers on GitHub. The repository provides a collection of Model Context Protocol servers designed to work smoothly on Windows systems. Its stated goal is connecting artificial intelligence tools to local resources and external services.","2026-11-03T05:03:24.810Z","news-windows-friendly-model-context-protocol-server-repository-published-on-github.webp",{"slug":71,"headline":72,"summary":73,"topic":48,"sourceName":13,"publishedAt":74,"coverImage":75},"trai-proposes-ai-governance-framework-for-telecom-operators","TRAI proposes AI governance framework for telecom operators","The Telecom Regulatory Authority of India (TRAI) has published draft artificial intelligence guidelines for telecommunications operators. The proposed framework mandates governance committees, structured risk assessments and compulsory reporting for serious AI-related incidents.","2026-11-02T05:02:17.136Z","news-trai-proposes-ai-governance-framework-for-telecom-operators.webp",1,true]