OpenClaw and NemoClaw AI agents exposed to critical vulnerabilities
OpenClaw and Nvidia's NemoClaw AI agent platforms were found to contain critical security vulnerabilities in April and May 2026, including improper access control and sandbox bypass flaws that could allow attackers to leak sensitive information. Chinese regulators issued warnings while Nvidia released patches; industry experts warned that such vulnerabilities in AI agents pose systemic risks to enterprise data security.
What happened
Between April 14 and 28, 2026, the China National Information Security Vulnerability Database (CNNVD) identified 111 vulnerabilities in OpenClaw, including 2 critical and 38 high-risk vulnerabilities covering access control errors, code issues, and path traversal flaws. The China National Internet Emergency Center and the Ministry of Industry and Information Technology (MIIT) issued continuous risk warnings about OpenClaw.
Nvidia issued a software security update in May 2026 to patch two information disclosure vulnerabilities in NemoClaw, its own OpenClaw-based development platform. CVE-2026-24222 allows remote attackers to send prompt injection requests causing improper access control, enabling agents to read and leak host environment variables when initializing sandbox environments (CVSS 8.6). CVE-2026-24231 exists in the validateEndpointUrl() SSRF protection component, allowing attackers to provide forged endpoint URLs to trigger server-side request forgery (CVSS 5.9 or 6.3).
The China National Computer Virus Emergency Response Center detected large quantities of counterfeit OpenClaw skill packages containing trojan viruses, posing serious risks to user data security and system stability.
Reported impact
Security experts and industry bodies including OWASP Foundation and Gartner identified AI agent security risks as a global industry consensus. Gartner predicted that by 2028, 25% of enterprise data breaches could be traced to agent misuse, involving both external malicious infiltration and internal unauthorized operations.
Index analysis
Industry experts noted that AI agents like OpenClaw face security challenges due to their 'intelligent emergence' bringing inherent uncertainty, causing collapse of identity systems and authorization planes. The vulnerability concentration suggests systematic security gaps in AI agent architecture affecting access control and sandbox isolation mechanisms.
Security researchers identified core threats including agent objective hijacking and tool misuse as outlined in OWASP Foundation's 'Top 10 Security Threats for Agent Applications.' The vulnerabilities in both OpenClaw and NemoClaw demonstrate risks in prompt injection handling and environment isolation that could be exploited during normal operation.
Governance controls
In April 2026, China's National Internet Information Office and four other departments jointly issued the 'Interim Measures for the Management of Anthropomorphic AI Interactive Services,' establishing inclusive and cautious, classified and graded supervision of AI services including security assessments, algorithm registration, and development of AI sandbox security service platforms. The MIIT and nine other departments jointly issued the 'Artificial Intelligence Science and Technology Ethics Review and Service Measures (Trial),' requiring robustness, controllability, trustworthiness, transparency and explainability, and accountability of model systems.
Industry standards for AI security are being accelerated, including 'AI Safety Governance Model Context Protocol Application Security Requirements' and planned standards such as 'AI Safety Capability Maturity Assessment Methods' and 'AI Application Security Classification and Grading Methods' under the National Cybersecurity Standardization Technical Committee.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.