Executive brief
PraisonAI, a framework for managing AI agents, contains a security misconfiguration in its prompt injection defense layer. By default, the system only blocks malicious inputs if they trigger three or more security checks simultaneously, allowing single-vector attacks to bypass protection. This could allow attackers to trick the AI into revealing sensitive system instructions, executing unauthorized commands, or manipulating financial transactions if the AI has access to payment tools.
Technical details
A vulnerability exists in the InjectionDefense class of PraisonAI due to an insecure default block threshold. The system is configured by default to only block inputs with a CRITICAL threat level, which requires three or more detection categories (such as instruction overrides, financial manipulation, or obfuscation) to trigger simultaneously. Consequently, single-vector prompt injection attacks that trigger a HIGH severity rating are logged but not blocked, allowing them to reach the underlying Large Language Model (LLM). Attackers can exploit this to perform system prompt extraction, unauthorized tool invocation, or session context exfiltration. The issue is resolved in version 4.6.78 by lowering the default block threshold to HIGH.
Affected products
- MervinPraison PraisonAI < 4.6.78
Timeline
- 2026-06-25: advisory: Vendor advisory published on GitHub
- 2026-07-11: disclosed: CVE published to NVD
- 2026-07-11: patched: Fix released in version 4.6.78