This journo found it serendipitous that big global AI players were calling for a pause on AI development just days before Zenith Live 2026 kicked off, with Anthropic’s proposal dominating headlines.
In Las Vegas, EITN seized the moment to ask Zscaler executives how they viewed the call for pause. To date, there have been a few high-profile calls for pause on building the most powerful AI systems, but no actual coordinated action has ever been implemented.
Since March 2023, the first open letter signed by prominent tech and AI figures had called for all AI labs to immediately stop training AI systems for at least six months. Since then, there have been grassroots movements across at least 13 countries that demanded governments to freeze developments of new frontier models until they can be properly evaluated and regulated.
Of course, this became the AI pause that never was. Until last June.
Anthropic’s announcement is the first time a leading frontier model lab has itself proposed a formal global mechanism to implement a moratorium on development of the very top tier of models.
AI itself is being used to help build more advanced AI, potentially leading to something called recursive self-improvement. Why is this risky?
Currently, humans still choose goals for AI systems, vet changes, and control deployment.
Without a human-the-loop, an AI system, in theory, can spot ways to make itself better, change its own design and training. It is improving its own ability to improve itself, creating a feedback loop where each upgrade makes the next upgrade faster or more powerful.
The model may get a fast take-off that outpaces existing safety controls and human oversight.
Zero trust was built for this moment
Against that backdrop, Zscaler execs like global CIO Sam Curry and EVP of AI Security Dhawal Sharma are dogged about one thing: the rules for discovering and mitigating security vulnerabilities still apply.
Zscaler is already treating AI as both a risk and a control plane. The company participates in OpenAI’s trusted‑access programmes to embed security‑tuned models directly into its own software development lifecycle: validating threat models at the design stage, assisting with secure code review and configuration checks, scanning dependencies to detect vulnerabilities earlier, and running automated black‑box testing on built artefacts. The development environment itself is being instrumented and defended by AI.
The company also hardens its own infrastructure and environment by having Claude Mythos Preview identify its flaws.
Regardless of any AI moratorium, a core priority for cyber defenders is to be ready for whatever vulnerabilities frontier models like Mythos and GPT5.5 Cyber will be able to find. This is Zscaler’s stance.
And Dhawal believes that there will be an inflection point whereby more software development will be done by more secure models. “At that point, some of these challenges will be behind us. But until then, we will have to take a very measured approach.”
We do think that zero trust was built for this moment.”
He argues that even if organisations have vulnerabilities that can’t be patched, “a zero trust architecture can hide those vulnerabilities from the internet. Attackers can’t reach you.”
What is possible now
Are frontier models showing signs they could slip away from human control and accelerate their own improvement? Anthropic’s report is that the latest frontier models are starting to help build the next generation of models.
Come what may on the policy front – pause, no pause, or something in between – if we can’t throttle the pace of model advancement, we can at least harden the architecture and infrastructure those models run on and ensure that every call and agent action traverses a zero trust fabric.