A newly disclosed three critical and high-severity vulnerabilities in Paperclip, an open-source control plane used to orchestrate autonomous "zero-human company" AI agents, have been exposed This article explores vulnerabilities paperclip. . This vulnerability is tracked as CVE-2026-41679 (CVSS 10.0, GHSA-68qg-g8mg-6pr7), affecting network-accessible authenticated deployments that run versions earlier than 2026.416.0 with default registration settings enabled.
An attacker can use a .paperclip.yaml bundle with a built-in process adapter to define an agent and execute arbitrary commands as Paperclip’s OS user, leading to full remote code execution including exposure of secrets, source code, and internal services. Unauthenticated users could access cross-tenant heartbeat-run issues, obtain agent-facing skill documentation for reconnaissance purposes, and read verbose /api/health responses that reveal deployment details, version numbers, and feature flags, which can help attackers identify vulnerable instances.
Because Paperclip accepted arbitrary Host headers and treated loopback traffic as trusted, the malicious page could silently import and start a process-based agent, executing commands with developer privileges without requiring credentials or cookies, simply by visiting a malicious site. Utilize in-browser data inspection from ANY.RUN for faster detection, investigation, and response capabilities, while maintaining comprehensive phishing visibility to enhance your SOC and reduce Mean Time To Repair (MTTR).
