AI security experts call for stronger deployment controls

NVIDIA has launched the Open Secure AI Alliance with a group of technology and security organisations. The initiative focuses on open approaches to artificial intelligence security.

The alliance brings together major industry players seeking a common approach to AI safety and governance. It emphasises open models and open infrastructure over closed systems.

Security specialists view the move as part of a broader shift in how organisations think about AI risk. Many now argue that the main weaknesses lie in how AI systems are deployed and controlled, rather than in whether the underlying models are open or proprietary.

Tim Hudson, President of OpenSSL Corporation and co-author of SSLeay, said the debate over model openness often distracts from more basic engineering work. He pointed to familiar concepts from traditional cybersecurity, such as identity and least privilege, as more reliable defences against misuse of AI systems.

"Those are engineering defects at the action boundary, and they are fixed with identity, least privilege, egress restriction, admission control, velocity limits and logs you can verify afterwards. Not one of them is fixed by a rule about who is permitted to possess a set of weights. If your security model depends on the adversary being unable to obtain a tool, you do not have a security model," Hudson said.

Hudson's comments reflect a growing view among security professionals that control at the point of use is more robust than control at the point of distribution. They argue that adversaries often gain access to tools and code, so system designers should assume that possibility and focus on monitoring, containment and response.

The launch of NVIDIA's alliance has also drawn support from vendors working on open-source AI infrastructure. They see the push for open and secure tooling as a way to standardise how enterprises run and measure AI workloads.

Debo Dutta, Chief AI Officer at Nutanix and an MLCommons Board Observer, said open foundations and shared benchmarks can give organisations a clearer view of how AI behaves in production environments.

"The foundation of secure enterprise AI must be built in the open," said Debo Dutta, Chief AI Officer at Nutanix and an MLCommons Board Observer. "By maintaining open-source infrastructure like KServe and Envoy AI Gateway, and co-creating rigorous benchmarks such as AILuminate, MedPerf, and MLPerf Storage, Nutanix is helping ensure that organisations have the transparent, scalable tools designed to deploy AI safely on their own terms."

Supporters of the alliance say open-source components and shared test suites make it easier to spot flaws in models and pipelines. They also say standard tools lower the barrier to audits and independent verification.

Vendors involved in AI security expect regulators and large customers to push for more demonstrable controls around access, logging and containment. They also expect greater scrutiny of how organisations combine open-weight models with sensitive data and business processes.

The debate is now shifting from whether models should be open to how organisations implement guardrails around their use. That includes enforcing identity, least privilege and admission control on every interaction with an AI system, along with verifiable logs that support post-incident investigations.

For Hudson, any framework that depends on keeping model weights out of adversaries' hands misses the core lesson from decades of internet security. He argues that security teams must assume tools will leak, then design AI systems that remain defensible under that assumption.