CISOs Battle AI Agent Risks: 78% Cite Security as Top Pain Point

Team8's annual CISO Village survey reveals a stark reality for enterprise security leaders: 71% of CISOs are now experimenting with or augmenting existing security tools using AI agent capabilities, yet 78% identify AI and agent security as their single biggest pain point—exactly double the second-ranked concern at 39%. The venture capital and private equity firm's flagship report, drawn from an invitation-only community of security executives at major global enterprises, highlights a widening gap between rapid AI adoption and the ability to govern it. The risk surface is expanding faster than the control layer, forcing security leaders to mobilize before they feel fully prepared.

Tim Brown, CISO at Team8, outlines two distinct challenges driving the anxiety. First, traditional cyber hygiene assumptions—MFA, firewalls, endpoint protection—are no longer sufficient against adversaries wielding AI-driven attack tools. As Brown told SecurityWeek: "The attacks and our attack surface have changed." The second, more pressing challenge is balancing AI-driven business enablement against the unintended consequences of over-privileged agentic AI. Employees across departments are now building agents using widely available coding platforms like Claude Code, Cursor, and Codex, ranging from simple email-summarization bots to complex sales-process automations that query enterprise data and touch critical network segments.

The deeper an agent reaches into enterprise systems, the higher the blast radius from poor instructions. Brown emphasizes a fundamental human limitation: people generally fail to articulate exactly what they want in plain text, and agents will dutifully attempt to execute ambiguous prompts against privileged systems—often with damaging results. Organizations deploying these tools should audit their current exposure using a privacy checkup to identify data leaks that over-privileged agents could inadvertently expose, while security teams evaluating AI tool vendors should run a WHOIS lookup and SSL/TLS checker on associated domains before granting access to internal environments.

The path forward, according to Brown and the CISO Village respondents, requires investing in new platforms, skills, and control mechanisms tailored to agentic AI—without throttling the business value these tools promise. With attackers already leveraging the same technologies, the window for getting that balance wrong is closing fast.