Researchers disclosed three (3) critical vulnerabilities in the open-source AI agent control platform Paperclip that could allow attackers to execute arbitrary commands or access sensitive information, with the most severe flaw, tracked as CVE-2026-41679 (CVSS 10), enabling unauthenticated remote code execution (RCE) on internet-accessible deployments configured with default registration settings. By abusing Paperclip’s import functionality and process adapter, attackers could register a new account, self-authorize a board credential, import a malicious AI agent, and execute commands with the privileges of the Paperclip server, potentially exposing source code, credentials, secrets, and internal services. A second critical flaw, tracked as GHSA-x8hx-rhr2-9rf7 (CVSS 9.6), leverages a DNS rebinding attack against Paperclip’s default local_trusted mode, allowing a malicious webpage to execute commands on a developer’s local machine without requiring authentication or stolen credentials. A third vulnerability, tracked as GHSA-xfqj-r5qw-8g4j (CVSS 8.3) stems from insufficient authorization checks across several API endpoints, potentially exposing deployment details, agent documentation, and cross-company data. Although no in-the-wild exploitation has been confirmed as of August 5, 2026, Rapid7 has released a public Metasploit module for CVE-2026-41679 and CISA’s SSVC assessment classifies proof-of-concept exploitation as available, significantly increasing the likelihood of future attacks. CTIX analysts strongly advise any affected organizations to upgrade to Paperclip v2026.416.0 (0.3.1) or later, which introduces stricter import authorization, hostname validation to prevent DNS rebinding, improved authentication and access controls, and additional safeguards against malicious agent imports.