{
  "id": 4029,
  "url": "https://arxiv.org/abs/2605.19722v1",
  "title": "Measuring Safety Alignment Effects in Autonomous Security Agents",
  "summary": "Do stock safety-aligned language models and their uncensored or abliterated derivatives behave differently when run as autonomous security agents? Single-turn refusal benchmarks cannot answer this question: security agents must inspect repositories, call tools, and produce vulnerability evidence inside authorized sandboxes. We present a trace-based benchmark of 30 local vulnerability-analysis tasks with fixed tools, deterministic success predicates, redaction rules, and grounding checks, and com",
  "authors": "Isaac David, Arthur Gervais",
  "category": "research",
  "topics": "safety-alignment,agents-autonomy",
  "orgs": null,
  "regions": null,
  "published_at": "2026-05-19T11:55:54.000Z",
  "fetched_at": "2026-07-14T16:30:41.583Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/4029",
  "original_url": "https://arxiv.org/abs/2605.19722v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}