CVE-2026-34938
PraisonAI’s multi‑agent system had a flaw that let attackers run any Python code on the host. By sending a specially crafted string subclass, the sandbox could be bypassed, giving full OS command execution. The issue was fixed in version 1.5.90.
An official fix is available, so this entry is kept for 30 days and then removed automatically.
Who is affected
PraisonAI praisonaiagents, any version before 1.5.90.
Real-world impact
An attacker who can send code to the execute_code() function could run arbitrary commands on the machine hosting the agents, potentially taking full control of the system.
Why this severity
The CVSS score of 10 reflects that the vulnerability is exploitable over the network, requires no user interaction, and gives an attacker complete control over confidentiality, integrity, and availability.
What to do about it
- 01Upgrade praisonaiagents to version 1.5.90 or later.
NVD-referenced vendor advisory
Timeline
- Apr 3, 2026 · Apr 3, 2026PublishedDisclosed and added to the National Vulnerability Database.
- Jul 24, 2026 · 11d agoAdvisory updatedThe NVD record was last revised.
- Jul 26, 2026 · 10d agoOfficial fix availableA vendor patch or mitigation now exists — see the remediation steps above.
How it’s attacked
References & advisories
- github.com/MervinPraison/PraisonAI/sec…exploitvendor advisory