A maximum-severity vulnerability in the open source AI agent platform Ruflo puts enterprise AI deployments at risk by letting attackers conduct various malicious activities from inside the orchestration framework. The flaw also can leave agents behaviorally compromised even after it’s been patched.
Researchers at Noma Security’s Noma Labs discovered the flaw, tracked as CVE-2026-59726, in Ruflo, formerly called Claude Flow and which hosts AI agent swarms for Codex and Claude Code, they revealed today. The vulnerability, which received the highest CVSS severity score of 10, allowed them to access the platform without logging in at all, according to Noma Labs.
“The weakness is a lack of authentication coupled with command execution capabilities, enabling complete control over the container and exposure of sensitive credentials,” according to details about the flaw posted on OpenCVE.
Memory Tampering Resists Fixes
Using their proof of concept (PoC), Noma Labs researchers accessed Ruflo via an unauthenticated model context protocol (MCP) bridge open to the network by default, demonstrating that a single HTTP request was enough to gain full remote code execution inside a Ruflo deployment, they said. Once inside, they could access the API keys Ruflo uses to talk to AI providers and read every user conversation stored on the platform.
Given that Ruflo is an AI agent host platform, the flaw could enable attackers to unleash a swarm of AI agents to do whatever they wanted.
Perhaps most troubling of all, researchers also found they could tamper with the AI’s own memory in ways that would influence its responses to future users long after the attacker had gone, they said. They did this by planting instructions that could steer Ruflo’s behavior even after they no longer were inside the system.
Noma Labs disclosed the vulnerability, which it nicknamed “RufRoot,” to Ruflo maintainers on June 30 and included a working PoC confirmed against a live default deployment. Ruflo responded within 24 hours to release a fix, defaulting the platform to a locked-down configuration with public exposure treated as an explicit opt-in requiring authentication, according to Noma, which independently verified the fix.
Still, merely patching software — the fix for traditional vulnerabilities in enterprise platforms — is no longer a guaranteed solution for a flaw that allows corruption of memory within an AI system, according to Noma Labs. Even with a patch, organization can still be running compromised agents. Indeed, this aspect of the vulnerability represents a new class of risk for enterprises deploying AI agents, experts say.
“This is not exploiting software flaws, it is memory poisoning,” observes John Gallagher, vice president at OT and IoT cyber hygiene firm Viakoo. Since the cyber threat impacts reasoning, the flaw represents “a critical turning point in cybersecurity, especially OT and IoT systems,” he says.
Defending Against Novel AI Threats
The Ruflo vulnerability discovery comes on the heels of an incident in which OpenAI models autonomously hacked the Hugging Face AI platform during benchmark testing, demonstrating that defenders are facing a new threat paradigm when it comes to protecting systems against the behavior of their own AI systems.
While the vulnerability’s lack of authentication is not new — and is, in fact, “one of the oldest failure patterns on the Internet” — what is new with AI agent abuse is “what sits behind that door,” observes Johan Edholm, security engineer and co-founder of app security testing firm Detectify.
“An agent platform concentrates everything an attacker wants in one place, with provider API keys, every stored conversation, shell access via its own tools, and a persistent memory the AI trusts,” he tells Dark Reading. “In other words, the bug class is old, but the blast radius is new.”
Indeed, if an attacker can corrupt an AI’s reasoning, organizations also need that same capability to use AI to fight AI, says Ram Varadarajan, CEO at cyber detection technology firm Acalvio. This could mean “corrupting the adversarial AI’s reasoning before it acts” and employing a bot-on-bot cyber defense, he says.
For now, Noma Labs advised, immediate remediation for those affected by CVE-2026-59726 includes treating AI provider credentials as compromised and rotating them, auditing the platform’s AI memory for tampering, and rebuilding containers from a clean image.
