CVE-2026-56074: MervinPraison PraisonAI tool approval cache bypass in praisonaiagents

CVE-2026-56074 · Severity: medium · CVSS 5.5 · Published 2026-06-18

PraisonAI, a framework for AI agents, contains a security flaw in how it handles user permission for running system commands. When a user approves a single safe command (like listing files), the system mistakenly remembers that approval for all future commands in that session. This allows an AI agent—potentially manipulated by a malicious prompt—to silently run dangerous commands to steal sensitive data like API keys and passwords without the user's knowledge.

Technologies: MervinPraison PraisonAI Agents, praisonaiagents (PyPI). Vendors: MervinPraison, PyPI.

Quick answers

Is CVE-2026-56074 exploited in the wild?
No. Junglewise Threat Intelligence has no report of CVE-2026-56074 being exploited in the wild.
How severe is CVE-2026-56074?
Medium severity, with a CVSS score of 5.5, published 18 June 2026.
Which products and versions are affected?
praisonaiagents (PyPI) ECOSYSTEM:introduced=0,fixed=4.5.128.
Is there a fix for CVE-2026-56074?
Yes. praisonaiagents (PyPI): upgrade to 4.5.128 or later. Junglewise finds every place you run it, drafts the upgrade and applies it once you approve.
How is CVE-2026-56074 attacked?
It is attacked with local access, in a software library.
Am I affected by CVE-2026-56074?
Check whether you run MervinPraison PraisonAI Agents or praisonaiagents (PyPI) at an affected version. Junglewise does that check for you around the clock: connect your code, cloud and domains, and it matches new vulnerabilities like this one against what you actually run and tells you when one reaches you.

How to fix

  • praisonaiagents (PyPI): upgrade to 4.5.128 or later
    pip install "praisonaiagents>=4.5.128"

Technical details

The vulnerability exists in the 'require_approval' decorator within the PraisonAI Agents approval system. The system caches approval status using only the tool name (e.g., 'execute_command') as the key, failing to include the specific invocation arguments in the cache lookup. Consequently, once a user provides manual consent for a benign command, all subsequent calls to that tool are automatically authorized for the remainder of the execution context. An attacker can leverage this via prompt injection to first request a harmless command and then silently execute shell commands to exfiltrate process environment variables (such as OPENAI_API_KEY) which are passed to subprocesses via 'os.environ.copy()'. The issue is addressed in version 1.5.128.

Affected products

  • MervinPraison PraisonAI (praisonaiagents) < 1.5.128

Identifiers

  • CVE-2026-56074
  • GHSA-ffp3-3562-8cv3
  • PYSEC-2026-2946

Timeline

  • 2026-04-09: advisory: GitHub Security Advisory published
  • 2026-06-18: disclosed: NVD publication date

References

This page as Markdown. Free to reuse under CC BY 4.0 with a link back.