Multiple security vulnerabilities have been identified in NeMo Agent Toolkit, TensorRT-LLM, and PhysX, affecting NeMo Agent Toolkit versions 1.0 through 1.8.0 and 1.4 through 1.6 for two of the issues. The flaws include missing authentication, SQL injection, untrusted functionality inclusion, out-of-bounds write, and server-side request forgery. The highest-severity issue allows remote unauthenticated attackers to compromise the toolkit, creating risk of code execution, privilege escalation, data tampering, denial of service, and information disclosure.
CVE-2026-65145 (CVSS 10.0 — Critical): A missing authentication flaw in NeMo Agent Toolkit allows a remote attacker to access critical functionality without logging in. The attacker can achieve code execution, privilege escalation, data tampering, denial of service, and information disclosure.[/subscribe_to_unlock_form]
Multiple security vulnerabilities have been identified in NeMo Agent Toolkit, TensorRT-LLM, and PhysX, affecting NeMo Agent Toolkit versions 1.0 through 1.8.0 and 1.4 through 1.6 for two of the issues. The flaws include missing authentication, SQL injection, untrusted functionality inclusion, out-of-bounds write, and server-side request forgery. The highest-severity issue allows remote unauthenticated attackers to compromise the toolkit, creating risk of code execution, privilege escalation, data tampering, denial of service, and information disclosure.
CVE-2026-65145 (CVSS 10.0 — Critical): A missing authentication flaw in NeMo Agent Toolkit allows a remote attacker to access critical functionality without logging in. The attacker can achieve code execution, privilege escalation, data tampering, denial of service, and information disclosure.[emaillocker id="1283"]
CVE-2026-65151 (CVSS 8.4 — High): A SQL injection flaw in NeMo Agent Toolkit allows an attacker who can supply crafted input to manipulate database queries. This can lead to unauthorized data access or modification.
CVE-2026-65119 (CVSS 8.4 — High): An inclusion of functionality from an untrusted control sphere flaw in TensorRT-LLM affects tokenizer loading. An attacker who can influence the loaded tokenizer content may achieve code execution.
CVE-2026-47486 (CVSS 7.8 — High): An out-of-bounds write flaw in PhysX can be triggered by processing crafted input. Successful exploitation may lead to memory corruption and code execution.
CVE-2026-65120 (CVSS 7.1 — High): A server-side request forgery flaw in TensorRT-LLM allows an attacker to cause the application to issue requests to unintended targets. This can expose internal services or data.
CVE-2026-24269 (CVSS 7.1 — High): A server-side request forgery flaw in TensorRT-LLM allows an attacker to cause the application to issue requests to unintended targets. This can expose internal services or data.
These vulnerabilities collectively present a severe risk to systems that expose the affected agent, model-serving, or physics components to untrusted networks. The greatest exposure belongs to deployments that allow remote access without strong authentication or that process untrusted input. Organizations should prioritize assessing exposure and treating these issues as urgent.
The following reports contain further technical details:
[/emaillocker]