{
  "id": 6214,
  "url": "https://arxiv.org/abs/2605.02914v1",
  "title": "When Safety Geometry Collapses: Fine-Tuning Vulnerabilities in Agentic Guard Models",
  "summary": "A guard model fine-tuned on entirely benign data can lose all safety alignment -- not through adversarial manipulation, but through standard domain specialization. We demonstrate this failure across three purpose-built safety classifiers -- LlamaGuard, WildGuard, and Granite Guardian -- deployed as protection layers in agentic AI pipelines, and show that it originates in the destruction of latent safety geometry: the structured harmful -- benign representational boundary that guides classificati",
  "authors": "Ismail Hossain, Sai Puppala, Jannatul Ferdaus, Md Jahangir Alam, Yoonpyo Lee, Syed Bahauddin Alam et al.",
  "category": "research",
  "topics": "safety-alignment,agents-autonomy",
  "orgs": null,
  "regions": null,
  "published_at": "2026-04-08T05:27:33.000Z",
  "fetched_at": "2026-07-14T16:32:20.055Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/6214",
  "original_url": "https://arxiv.org/abs/2605.02914v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}