We scanned Common Crawl - a massive dataset used to train LLMs like DeepSeek - and found ~12,000 hardcoded live API keys and passwords. This highlights a growing issue: LLMs trained on insecure code may inadvertently generate unsafe outputs.