Study: LLMs reproduce real leaked API keys from training data

Researchers recovered strings matching real DeepSeek and Google API keys from Qwen3-27B, GPT-Codex 5.3 and Claude Haiku 4.5 using output-only API access, with no access to weights or token probabilities. Every recovered key had a matching public occurrence on GitHub. GitGuardian separately found 28.65 million new hardcoded secrets in public GitHub commits in 2025.

• Qwen3-27B returned two DeepSeek keys and one Google key; GPT-Codex 5.3 and Claude Haiku 4.5 returned DeepSeek keys
• Every recovered credential had a matching public occurrence on GitHub
• 28.65 million new hardcoded secrets appeared in public GitHub commits in 2025
• Claude Code-assisted commits leaked secrets at 3.2% versus a 1.5% baseline