Presented by Snowflake
In the realm of enterprise security, history often reflects a pattern of increasing complexity. New threats lead to additional controls, which in turn encourage users to circumvent the very systems intended to protect them.
Throughout my career, I’ve observed that the main obstacle to security adoption is not a lack of concern for safety but rather the perception that secure methods are more cumbersome than insecure alternatives.
This insight is particularly pertinent in the era of AI.
AI broadens the potential attack surface and enhances the capabilities of attackers, making the simplification of security measures increasingly vital. Controls that prove inconvenient are eventually disregarded, prompting workarounds. The solution lies in ensuring that the secure route is the most straightforward one.
Security is most effective when unobtrusive
Security gains traction when it’s easier to comply than to bypass. When two-factor authentication was first scaled in the industry, the challenge was not the security itself but the disruption it caused, requiring users to pause their tasks, use a phone, initiate a VPN, and enter codes, thereby interrupting their workflow.
Adoption was driven by simplicity rather than policy or training. With innovations like fingerprint and facial recognition, users now embrace it effortlessly.
This principle also guided browser developers to make security more user-friendly. Instead of relying on users to scrutinize URLs, modern browsers now visibly mark non-HTTPS sites as insecure, steering users toward safer practices by default. Security strengthened as the safer path became more intuitive and accessible.
Complexity in AI systems
Agent permissions illustrate complexity within AI systems. Over time, employees acquire a myriad of permissions through various assignments, system accesses, and roles that are often not updated post-team changes. Humans inherently recognize relevant access, even if systems don’t enforce it.
Conversely, agents lack this discernment. An agent tasked with a problem may explore every option available, even when only a fraction is necessary, inadvertently expanding the potential attack surface.
Although involving a human to approve critical actions seems prudent, agents may request approval for technical actions without sufficient context, leading to approvals that maintain workflow at the expense of oversight.
A more effective approach is an intent-based permission model. Agents should receive credentials limited to specific tasks, with automatic expiration post-completion. Industry standards like OAuth are evolving to accommodate agentic AI, allowing agents to carry task-specific identities rather than full user permission sets.
Streamlining AI security
Improving ease of use begins with transparency. The initial step is understanding agent activities, such as connection points, data access, and permissions utilized.
Enterprises often find their visibility limited to around 80%. The remaining 20% poses significant risk, with AI potentially identifying these gaps faster than humans. Initiate with monitoring, even without immediate enforcement plans. Leverage AI to analyze findings, prioritize high-risk behaviors, and systematically address them.
In terms of identity, shift towards workload identity where feasible. The traditional method of creating service accounts and distributing static keys is both fragile and difficult to audit. Modern cloud solutions offer a more robust system, establishing a workload’s identity at deployment, eliminating static key distribution, reducing management burdens, and minimizing attack surfaces.
For agents, avoid granting broad permissions under the assumption that human oversight will prevent issues. Limit access to necessary tasks, ensuring permissions expire upon task completion. MCP gateways are becoming a practical solution for teams managing multiple agent-tool connections, enabling centralized governance rather than tool-specific rules. Keep human oversight for significant actions, especially those with substantial potential impact.
Accelerating risk pace
In the AI age, the time between exposure and exploitation is shrinking rapidly, sometimes from days to mere minutes. CrowdStrike’s 2026 Global Threat Report indicates a 65% year-over-year increase in attacker breakout speed. As AI autonomously detects vulnerabilities, security teams relying solely on manual responses risk falling behind.
The solution remains unchanged. Security measures that introduce friction will be circumvented. Effective security is embedded within the architecture, enforced by default and unobtrusive in operation. Although AI elevates the stakes, the fundamental principle persists: security is only effective when the secure path is also the simplest.
Mayank Upadhyay is Chief Security & Trust Officer at Snowflake.
Sponsored articles are content produced by a company that is either paying for the post or has a business relationship with VentureBeat, and they’re always clearly marked. For more information, contact sales@venturebeat.com.
