Research
Practical Secrets Extraction against Black-box LLMs
First reported · Discovered arxiv.org
Page published
Earliest dated coverage: 19 Oct 2026 · First observed: 27 Oct 2026 · Latest dated coverage: 19 Oct 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
Black-box credential extraction shows that leaked secrets can persist in deployed LLM outputs, reinforcing the need to revoke exposed keys and prevent coding agents from ingesting fresh credentials.
Researchers behind “Practical Secrets Extraction against Black-box LLMs” report recovering API-key strings from Qwen3-27B, GPT-Codex 5.3, and Claude Haiku 4.5 using output-only API access and a locally distilled proxy model. Every recovered credential matched a public GitHub occurrence, although the researchers did not test whether the keys remained valid.