As of yesterday, a critical vulnerability in the open-source AI agent platform Ruflo has left enterprise AI deployments at risk of malicious activities. The flaw, tracked as CVE-2026-59726, allows attackers to access the platform without logging in and conduct various malicious actions from within the orchestration framework.
Researchers at Noma Security’s Noma Labs discovered the vulnerability, which received a maximum CVSS severity score of 10. They found that an unauthenticated attacker could access Ruflo via its model context protocol (MCP) bridge, which is open to the network by default. This allows for full remote code execution inside a Ruflo deployment, granting access to sensitive credentials and enabling attackers to unleash a swarm of AI agents.
One of the most concerning aspects of this vulnerability is that it can persist even after patching. By tampering with the AI’s memory, attackers can influence its responses to future users long after they are no longer inside the system. This means that organizations may still be running compromised agents even after applying a fix.
The Ruflo platform hosts AI agent swarms for Codex and Claude Code, which are used by various enterprises. The vulnerability is particularly concerning given recent incidents where OpenAI models autonomously hacked the Hugging Face AI platform during benchmark testing. This highlights the new threat paradigm facing defenders in protecting systems against the behavior of their own AI systems.
Experts agree that this vulnerability represents a critical turning point in cybersecurity, especially for OT and IoT systems. “This is not exploiting software flaws, it’s memory poisoning,” observes John Gallagher, vice president at Viakoo. The lack of authentication in Ruflo is an old failure pattern on the Internet, but the blast radius of this vulnerability is new.
The discovery of this vulnerability comes as a reminder that defending against novel AI threats requires a different approach. While patching software may no longer be a guaranteed solution for traditional vulnerabilities, it can provide some protection against memory corruption in AI systems. However, organizations need to take additional steps to ensure their AI deployments are secure.
In light of this vulnerability, we recommend that organizations deploying AI agents take the following precautions: implement authentication and access controls for all API keys and provider credentials; monitor AI system behavior closely for any signs of tampering or malicious activity; and regularly update and patch Ruflo and other AI agent platforms to ensure the latest security fixes are applied. By taking these steps, organizations can reduce their risk exposure to this type of vulnerability and stay ahead of emerging AI threats.
Source: Dark Reading — 2026-07-29