{"data":{"id":"6f73e2e4-31ba-476c-8991-5ae32188a78f","title":"Follow the Gradient: Tracing Safety Degradation Back to Training Data","summary":"This paper presents SimpleLoRIF, a gradient embedding method that flags instruction-tuning examples whose gradients align with unsafe behavior, even when the examples look benign. Applied to Tulu 3, removing the 20% most aligned examples to produce SafeTulu reduced GCG attack success rate by up to 22% across Qwen, Llama, and Olmo, with capability largely preserved. On emergent misalignment, the authors report AUC as high as 0.865 in the gradient space of misaligned models.","solution":"The source describes removing the 20% of training examples most gradient-aligned with unsafe responses, producing the SafeTulu dataset, and releases SafeTulu along with a Hugging Face Space that lets researchers customize the removal fraction or filter subsets.","labels":["research","safety"],"sourceUrl":"https://arxiv.org/abs/2610.37481v1","publishedAt":"2026-10-27T16:08:08.000Z","cveId":null,"cweIds":null,"cvssScore":null,"cvssSeverity":null,"severity":"info","attackType":[],"issueType":"research","affectedPackages":null,"affectedPackageNames":null,"affectedPackageRefs":null,"affectedVendors":["Meta","HuggingFace"],"affectedVendorsRaw":["Qwen","Llama","Olmo","Tulu 3","SafeTulu","Hugging Face"],"classifierModel":"claude-haiku-5-5","classifierPromptVersion":"v4","summaryPromptVersion":"v2","headline":null,"headlinePromptVersion":null,"cvssVector":null,"attackVector":null,"attackComplexity":null,"privilegesRequired":null,"userInteraction":null,"exploitMaturity":null,"epssScore":null,"epssCheckedAt":null,"kevDateAdded":null,"advisoryAliases":null,"affectedPackagesSource":null,"affectedPackagesCheckedAt":null,"patchAvailable":null,"disclosureDate":"2026-10-27T16:08:08.000Z","capecIds":null,"crossRefCount":0,"attackSophistication":"moderate","impactType":["safety","integrity"],"aiComponentTargeted":"training_data","llmSpecific":true,"classifierConfidence":0.93,"researchCategory":"preprint","atlasIds":null}}