Building Trust in Autonomous AI
We're excited to introduce our new engineering podcast, hosted by David Huish, Head of Engineering at Aspiron Search. In each episode, we sit down with engineering experts to explore the latest developments, challenges and conversations shaping cybersecurity.
For our latest episode, David was joined by Sanket Vasa, Principal Software Engineering Manager at Microsoft, who works on governance and security for agents within Copilot Studio.
With almost 20 years at Microsoft, Sanket has seen plenty of changes in software engineering. Now, he's helping tackle one of the biggest challenges facing businesses adopting agentic AI: how do we make sure we can trust it?
From AI transforming engineering teams to the importance of identity, permissions and governance, we explore what it takes to build and secure the next generation of autonomous AI.
AI is giving engineering teams more room to build
For Sanket, one of the biggest changes AI has brought to his role is the time it gives back.
By using coding agents to handle repetitive tasks, project management and planning, he's been able to spend more time building solutions, developing features and working directly with engineers.
It's not just about getting things done faster. It's about changing how engineers spend their time and giving them more room to focus on the work that makes a difference.
Why agent security needs the right boundaries
As agents become more capable, controlling what they can access is increasingly important.
At Microsoft, Copilot Studio gives administrators control over which models agents can use, which systems they can connect to and how widely they can be shared.
For example, an agent might be restricted from accessing certain third-party connectors if there's a risk of company data being exposed.
These policies are enforced both when agents are being built and while they're running.
Sanket also highlighted the importance of ensuring employees only access information relevant to their roles. A HR employee shouldn't automatically have access to finance systems simply because an AI agent can connect to them.
Why identity is becoming so important
One of the biggest shifts Sanket has noticed is the move away from giving agents the same permissions as the person using them.
Initially, many agents operated using the permissions of the person interacting with them, potentially giving them access to more information than they actually needed.
However, agents can hallucinate, fall victim to prompt injection attacks and behave unpredictably.
The industry is now moving towards dedicated agent identities, isolated environments and least-privilege access.
In simple terms, agents should only have access to what they actually need to do their job.
Will AI replace software engineers?
It's difficult to have a conversation about AI without the question of jobs coming up.
But Sanket is optimistic about what agentic AI means for the future of engineering.
He sees coding agents as the next step in software development, allowing engineers to spend less time writing code manually and more time identifying problems, reviewing solutions and focusing on customer needs.
He predicts that within the next one to two years, the vast majority of code could be written, reviewed and validated by agents, with humans overseeing the process.
Rather than seeing AI as something that takes opportunities away, he sees it as a way for people to achieve more.
Building trust is the next step
For Sanket, one of the biggest barriers to wider enterprise AI adoption isn't necessarily the technology itself. It's having the right security and governance in place.
Businesses need confidence that their agents can operate safely, access the right information and respect the boundaries put in place.
His advice for anyone looking to build a career in AI security? Get the fundamentals right. Identity management, permissions and defence in depth are just as important for agents as they are for traditional systems.
And don't be afraid to get hands-on. Build agents, experiment with their capabilities and see what happens when you give them too much access.
The opportunity with autonomous AI is huge. But building trust in it is what will allow businesses to make the most of it.
Check out the full podcast here!
Frequently Asked Questions
1. What is agentic AI?
Agentic AI refers to AI systems that can independently carry out tasks, make decisions and interact with tools or business systems to achieve a particular goal, rather than simply responding to prompts.
2. Why is security important for autonomous AI agents?
AI agents can access sensitive information, interact with external systems and take actions independently. Without the right permissions and controls, they can introduce risks such as data exposure, prompt injection and unauthorised access.
3. What is least-privilege access in AI security?
Least-privilege access means giving an AI agent only the permissions it needs to complete its specific tasks, rather than granting it broad access to systems and information.
4. Will AI agents replace software engineers?
AI agents are expected to automate more software development tasks, including writing and reviewing code. However, human oversight will remain important for quality, security, decision-making and understanding business requirements.
5. What is holding back enterprise AI adoption?
A lack of robust security controls, governance and identity management can make businesses hesitant to move AI projects from experimentation into production.