Maximum severity vulnerability in Ruflo AI platform allows memory tampering

A maximum-severity vulnerability in the open-source AI agent platform Ruflo puts enterprise AI deployments at risk by allowing attackers to conduct malicious activities from within the orchestration framework. The flaw, tracked as CVE-2026-59726, also can leave agents behaviorally compromised even after patching, with further coverage provided by Dark Reading.

Researchers from Noma Security discovered the flaw in Ruflo, formerly Claude Flow, which hosts AI agent swarms for Codex and Claude Code. The vulnerability, with a CVSS score of 10, allows unauthenticated attackers to gain full remote code execution within a Ruflo deployment. Attackers can access provider API keys, read stored user conversations, and tamper with the AI's memory, influencing future responses. This memory tampering, described as "memory poisoning," can persist even after patching, representing a new class of risk for enterprises.

Noma Labs disclosed the vulnerability on June 30, and Ruflo released a fix within 24 hours, defaulting the platform to a locked-down configuration. However, experts note that patching alone may not be sufficient due to the potential for persistent memory corruption. Immediate remediation advised includes rotating AI provider credentials, auditing AI memory for tampering, and rebuilding containers from a clean image.

Source: Dark Reading