NVIDIA called on the software industry to treat artificial intelligence security as an engineering discipline with enforceable runtime limits, designated owners, and reproducible test evidence across every layer of the agent stack.
Modern AI systems combine models, workflow harnesses, and execution runtimes, NVIDIA stated. The underlying runtime environment must enforce network, file, and process limits independently of an agent's internal reasoning. Under this setup, an agent updating a customer record cannot authorize an export of that data, even when prompted by malicious instructions embedded in an attached file. Protected logs must capture the tool call, authorization choice, and destination.
To enforce these boundaries, NVIDIA released OpenShell, an open-source sandboxed runtime that restricts how agents access files, networks, and system resources. Partners in the Open Secure AI Alliance are already using the technology. Cisco is building a governance layer called DefenseClaw, while JFrog integrates with OpenShell to scan agent skills and enforce access rules on external dependencies.
NVIDIA emphasized that engineering teams must appoint named owners who review test evidence before clearing agents for production. Testing must verify that controls block credential theft, unauthorized exports, and attempts to tamper with logging. NVIDIA pointed to CrowdStrike’s SafeMind for automated attack simulations and Palo Alto Networks Prisma AIRS for continuous red teaming.
For vulnerability inspection, NVIDIA highlighted Capital One’s VulnHunter for code analysis and ReversingLabs’ Spectra Assure for detecting tampered software packages. Open models allow defense teams to reproduce system failures and validate fixes locally without sending sensitive forensic evidence outside their own infrastructure.
