When working with Generative AI (GenAI) day in and day out, you see the promise and experience the difficulty of an entirely new business paradigm. GenAI brings real, transformative, enterprise scale power to knowledge work – it is the coming of the second industrial revolution. Yet, organizations shouldn’t assume agents will behave benevolently when deployed into workflows. Like an errant knight, agents will follow whatever quest they happen to receive through user prompting – malicious or not. We would be foolhardy to trust agents, and we must design our infrastructure accordingly around the principle of zero trust.
But what is zero trust in the Agentic age? A case can be made that zero-trust equates to information security. Imagine a future where agents are given access to all data domains within an org and that those agents can talk to each other, just like people, to accomplish their various tasks. How do you monitor and govern the flow of information in that system? How do you ensure that sensitive information remains secure as it crosses compartmentalized permission boundaries platforms or borders? Furthermore, the ephemeral and scalable nature of agents dramatically increases the attack surface area for data breaches. If an agent is jailbroken, or if a prompt injection succeeds, how far can it reach? What data could it exfiltrate? What systems could it impact? These risks should drive architecture and access control decisions. If you build security at the data layer, then your data and information is secure by default and you can achieve zero trust.
YOU GAVE YOUR AGENTS POWER NOW GIVE THEM PRINCIPLES.
Agentic systems are expected to have the same flexibility as a regular human employee with only a fraction of the human context. Each agent lacks a strong world model of security risk. An employee understands the consequences of misuse – an agent does not. The nuance of security culture is tough to capture in the prompts of an Agentic system, and guardrails can only go so far in curtailing errant behavior. Like Don Quixote, your Agents will most definitely tilt at windmills if given the opportunity.
The aforementioned governance and compliance are often seen as friction. However, in security conscious organizations they are strategic enablers that allow stakeholders to trust in rapid development. When agents are given the authority to access sensitive data, make decisions, and act on behalf of people, we must also ensure that they abide by strong security principles. If we can’t rely on the internal security world model of agents, it follows that we must build using design patterns that are secure by default, with limited permission scopes, layers of authority, and human in the loop checks. Without those elements, agents expose the enterprise to risks that are not easy to quantify.
A corollary to building with secure architecture is the need for adversarial testing. Give your red-team access to the agent without constraints and try all manner of prompt injection. Give them a compromised system and let them attempt to expand its reach. What can they extract? How many systems can they access?
For companies with sovereignty requirements, data that must stay within a country, models that must run under your control—the only realistic path is unified and centralized data control over infrastructure, compute, data, and pipelines. If you relinquish any one component, you jeopardize the guarantee that everything remains internal and safe.
PROTEGRITY SECURE DATA PLATFORM ENABLED A GLOBAL TOP 5 BANK PRESERVE DATA SOVEREIGNTY FOR CUSTOMERS IN 130 COUNTRIES.
Developing complex and powerful Agentic systems is only the first step in realizing the second industrial revolution. The follow-on challenge is building the trust required to rapidly move Agents into production. Data security is a powerful tool to control risk, build trust, and accelerate GenAI adoption.
Protegrity Accelerates AI Adoption
Protegrity protects the sensitive data fueling your AI and LLM initiatives, whether training models or powering copilots, ensure responsible AI with privacy, security, and guardrails built in.