Security used to be the team that said no. Arshad Ahmad, Director of Information Security, Architecture and Engineering at ZS, argues that in the age of AI it has to be the team that helps you say yes safely.
On The Builders Club Startup Founders Podcast, he shares what he has learned securing data in pharma, fintech and consulting, and what leaders should do about AI security now.
How is AI changing data security? Arshad Ahmad’s core thesis
Arshad’s core view is that security should enable innovation rather than block it. If security only blocks, people build shadow IT, so the better path is to understand how the business works and wrap defence in depth around it.
On AI specifically, he says to treat an AI agent like an intern: train it, test it in a sandbox, add guardrails and keep a human in the loop before it gets real authority.
- Enable, don’t block. Work with scientists and business teams so safe options are also the easy ones.
- Data quality comes first. The model matters less than whether the data feeding it is correct and clean.
- Check both sides. Validate prompts going in and answers coming out with an AI runtime layer.
- Attackers use AI too. Phishing is now targeted, well written and automated.
Why should security enable the business instead of blocking it?
Arshad describes pharma teams as highly technical and creative. If security keeps blocking them, they find workarounds, so he prefers to sit with them and design the controls around their work.
He gives a lab example: tissue scanners that produced very large image files, with runs lasting days. A network interruption meant starting over, so reliable infrastructure was part of security, not separate from it.
- Put older lab equipment behind network segmentation such as a DMZ.
- Apply a zero trust architecture so work continues but access is verified.
- Build reliability into the design, since failed runs and clinical trials are costly.
What does an enterprise security review of a new project look like?
Arshad walks through how ZS handles a new client application. The process starts before the build and follows it through development.
- Governance, risk, compliance and legal teams assess the project, because data rules differ across the US, Europe, Japan and China.
- The information security team reviews architecture, infrastructure, application security, access controls and IP protection.
- During development, code is scanned and tested through a secure development lifecycle.
- The same process repeats for every new application.
How are pharma companies using AI today?
He points to three areas. In drug discovery, AI predicts which combinations are worth trying, so the wet lab tests fewer candidates in a dry-lab-first approach.
In clinical trials, AI can scan documents for gaps and help find the right patients, and commercial teams can query data in natural language instead of waiting for a data scientist.
A caution from an AI drug-selection experiment
Arshad recalls a project at his previous company where models were fed limited clinical trial data and asked to suggest medicines for symptoms. Accuracy started very low and reached roughly 70 to 80 percent after months of added retrieval work.
That is far short of what medicine requires, and the model sometimes picked a drug from a failed trial. His conclusion was that a human must stay in the loop.
How do you secure AI with runtime guardrails?
Arshad says input prompts and output responses are the two control points. An AI runtime layer, effectively a proxy, checks whether a question is allowed before it reaches the model, then checks whether the answer is something the user may see.
His HR example: an employee asking for another person’s salary should be blocked on the way in, and if it slips through, blocked again on the way out.
- Input validation. Is this user authorised to ask this?
- Output validation. Is this user entitled to the data in the answer?
- Data ownership. Under rules like GDPR, the data owner stays responsible for what is fed in.
What new threats does AI bring, and how should leaders respond?
Attackers use AI too. Phishing emails no longer have telltale grammar mistakes, and they carry real context from company news or reports, aimed at specific teams, with bots running around the clock.
His answer on defence is AI for security operations, with human oversight. He compares an AI agent to an intern who is not handed critical credentials on day one.
- Zero trust. Insider and compromised-account risk makes trust-but-verify essential.
- Review vendor contracts. Tools that add AI after signing may change how your data is used.
- Prepare for post-quantum cryptography. Start inventorying assets and exploring new methods.
- Train users. Awareness of AI-specific threats still matters.
Will AI replace jobs, and how should families approach AI?
Arshad does not expect AI to replace jobs soon. He compares it to Office 365, which cut the staff needed to run mail servers but created new work, and expects companies to ship more products with the same people.
He also advises guiding children to use AI as a tutor, verify answers across sources and understand that models can be fed wrong information.
Further reading and sources
- ZS official website
- Zero trust security model on Wikipedia
- General Data Protection Regulation on Wikipedia
- Post-quantum cryptography on Wikipedia
Frequently asked questions
What is AI runtime security?
It is a layer that sits between users and an AI model and checks traffic in both directions. Prompts are checked for authorisation before they reach the model, and answers are checked before they are shown. Arshad Ahmad says this input and output validation is an area many organisations, including ZS, are exploring.
How should companies roll out AI agents safely in security?
Arshad suggests treating the agent like an intern. Train it, test it in a sandbox or on low-risk tasks, add guardrails, and only expand its authority as trust builds, keeping a human in the loop for critical actions.
What should leaders check before adopting AI features in existing software?
Review your vendor contracts. A tool bought before it had AI may now process your data in new ways, so confirm the agreement covers how that data can be used. Pair this with zero trust access and user awareness training.
Conclusion
Arshad Ahmad’s advice is to make security a partner to the business: design controls around real workflows, validate AI inputs and outputs, and keep humans in charge of critical decisions.
Start this quarter by reviewing your AI-enabled vendor contracts and deciding where an AI agent would start as an intern.
Want more conversations like this? Join The Builders Club, our community of 35,000+ founders and CXOs, and explore more episodes of the Startup Founders Podcast.
