Executive brief
PraisonAI is a multi-agent framework that can integrate with email services. When an AI agent is configured with email credentials, malicious LLM outputs can exploit unsafe string interpolation in the email tool to inject commands into IMAP search queries. An attacker controlling the agent's output (e.g., through prompt injection) can read, modify, or delete mailbox contents, or disrupt email connections.
Technical details
The vulnerability is an IMAP injection flaw in email_tools.py where LLM-controlled values (from_addr, subject, query) are interpolated directly into quoted IMAP SEARCH criteria without proper escaping. Special characters—quotes, backslashes, newlines, and nulls—can break out of the quoted string and inject arbitrary IMAP commands. The attack requires an AI agent with configured email credentials to be exposed to untrusted input or prompt injection. An attacker can alter IMAP operations to access, modify, or delete messages, or corrupt the connection. The fix in praisonaiagents 1.6.59 implements proper IMAP protocol escaping.
Affected products
- Mervin Praison PraisonAI Agents before 1.6.59
Timeline
- 2026-09-14: disclosed
- 2026-06-17: patched: Fix released in praisonaiagents 1.6.59