Deepfake Detection for Agent Identity: Lessons from 1.5M Leaked Keys

Wiz researchers found that Moltbook, a viral social network where AI agents post and interact, left a database publicly readable with no row-level security. Inside: roughly 1.5 million API keys in plaintext for OpenAI, Anthropic, AWS, GitHub and Google Cloud, plus private agent-to-agent messages and owner email addresses. The same misconfiguration even allowed writes to any post.

Keys are identity

An API key is not a password you memorise. It is a machine identity. Whoever holds the key is the agent, as far as every connected service is concerned. With these keys, anyone could impersonate arbitrary agents, read their private conversations, or reach into the cloud accounts behind them. The keys stayed valid until each owner rotated them, which for 1.5 million keys means many never got rotated at all. The deeper problem is that AI agents are being handed the keys to production systems with the security practices of a weekend hackathon. Every agent with a leaked key is a trusted insider working for whoever found the database.

The uncomfortable question is how anyone verifies that an "agent" online is who it claims to be. Deepfake detection started with fake faces, but the underlying problem is identical: proving the source of digital content. Ai-powered deepfake detection platforms are really in the provenance business. When agents post, trade and message with real credentials, knowing which outputs came from which agent is the whole game.