Executive brief
PraisonAIAgents is a framework for building AI agents. A security flaw in its file-reading tool allows an AI agent—potentially manipulated by malicious input—to read any file on the host system, such as passwords or private keys, without the user's permission. This could lead to the theft of sensitive credentials and private data.
Technical details
The `read_skill_file` function in `skill_tools.py` fails to validate that the `skill_path` parameter remains within the intended workspace directory. While it contains a check to prevent directory traversal within the provided path, the base path itself is unrestricted and can be set to absolute system directories (e.g., `/etc` or `/root/.ssh`). Additionally, unlike other dangerous tools in the suite, this function was not protected by the `@require_approval` decorator or included in the `DEFAULT_DANGEROUS_TOOLS` registry. An attacker leveraging prompt injection can trigger this tool to read and return sensitive files like `/etc/shadow` or environment variables containing API keys. The vulnerability is addressed in version 1.5.128.
Affected products
- MervinPraison praisonaiagents < 1.5.128
Timeline
- 2026-04-09: disclosed
- 2026-04-10: advisory: GitHub Advisory GHSA-grrg-5cg9-58pf published
- 2026-04-10: patched: Version 1.5.128 released