The corporate landscape is currently undergoing a seismic shift. According to the latest "State of AI in the Enterprise" report from cloud content management leader Box, 83% of enterprises have now deployed AI agents into their operational workflows. This rapid adoption signifies a move beyond simple, reactive chatbots toward proactive, autonomous "agentic" systems capable of executing complex, multi-step tasks with minimal human intervention.
While these tools promise unprecedented levels of productivity and operational efficiency, they have simultaneously introduced a new frontier of security vulnerabilities. As headlines across the tech sector highlight the risks of "shadow AI" and compromised agentic workflows, security leaders are faced with a daunting mandate: how to innovate at speed without compromising the integrity of the enterprise perimeter.
In a recent episode of the ITPro Podcast, Box Chief Information Security Officer (CISO) Heather Ceylan joined hosts Ross Kelly and Bobby Hellard to dissect this dichotomy. The discussion offered a blueprint for how organizations can navigate the complexities of agentic AI while maintaining a robust security posture.
Main Facts: The Rise of Autonomous Agents
At the heart of the current AI transition is the transition from "generative" AI—which creates content—to "agentic" AI, which performs actions. Unlike static large language models (LLMs) that respond to prompts, agents are designed to interface with software, navigate systems, and resolve issues autonomously.
The 83% adoption rate cited by Box suggests that the experimental phase of AI is effectively over. Enterprises are moving these agents into production, integrating them into security operations centers (SOCs), software development lifecycles, and customer service pipelines. However, the speed of this transition has outpaced the development of standard governance frameworks, leaving many CISOs to play catch-up as they manage the risks of these autonomous digital workers.
Chronology of the Agentic Evolution
The trajectory of AI adoption has been anything but linear.
- Phase 1: The Chatbot Era (2022–Early 2023): Organizations experimented with off-the-shelf generative AI to draft emails and summarize documents. These were largely isolated from internal databases and carried minimal systemic risk.
- Phase 2: Integration (Late 2023): Companies began plugging AI models into their internal data stacks. This opened the door for RAG (Retrieval-Augmented Generation) but also introduced concerns regarding data privacy and unauthorized access.
- Phase 3: The Agentic Transition (2024–Present): The current era is defined by autonomy. Agents are no longer just "talking"; they are "acting." They have the capability to write code, execute patches, and query sensitive log files. This shift has necessitated a fundamental rethinking of the "Human-in-the-Loop" (HITL) model.
Supporting Data: Security and Efficiency
The promise of agentic AI lies in its ability to handle the "drudgery" of enterprise operations. Ceylan notes that for security teams, agents provide the long-awaited capacity to outpace attackers.
Operational Gains in the SOC
The Security Operations Center (SOC) is perhaps the most significant beneficiary of agentic intervention. According to Ceylan, human analysts are often overwhelmed by the volume of alerts, many of which are false positives. By delegating the triage, enrichment, and log correlation to agents, teams can reduce their mean-time-to-respond (MTTR) significantly.
However, Ceylan is careful to qualify this: "There’s still human judgment in the end in terms of what gets escalated to be an incident and what doesn’t." The agent acts as a force multiplier for the human analyst, not a replacement for the final decision-making authority.
Automated Remediation in Development
Perhaps the most ambitious deployment of agents at Box is within the software development lifecycle (SDLC). By embedding agents into the architecture review process, Box is automating the identification and—crucially—the enforcement of security fixes. This represents a shift from "detect and report" to "detect and resolve," a leap in capability that traditional static analysis tools could never achieve.
Official Perspectives: The Multi-Model Strategy
One of the most profound takeaways from the conversation with Ceylan is the departure from "vendor lock-in" regarding AI models. A year ago, many enterprises were searching for the "single best" model. Today, the strategy has shifted toward a multi-model approach.
"We’re not in a place where most of the work we do, we can’t be reliant on a single model," Ceylan explained. By utilizing a diverse array of models—whether for vulnerability discovery or code review—enterprises can hedge against the inherent biases and limitations of any single AI provider. This approach not only improves accuracy but also adds a layer of resilience; if one model is compromised or exhibits "hallucinations" in a specific domain, other models can serve as verification agents.
Implications for the Enterprise
The integration of agentic AI brings significant, often overlooked, implications for organizational structure and risk management.
The Problem of "Security Fatigue"
Ceylan acknowledges that the sheer pace of AI advancement is causing burnout among security professionals. "Things are changing quickly. Sometimes it feels like you take two steps forward and then you read something in the news and you’re like, ‘Oh my gosh, we need to rethink everything.’"
This environment requires a shift in the security mindset from "set-it-and-forget-it" to "continuous adaptation." CISOs must now cultivate teams that are not only skilled in traditional network security but are also proficient in "AI Red Teaming" and model governance.
The Governance Gap
As agents become more autonomous, they inherently accrue more permissions. If an agent has the ability to "enforce fixes," it must have write-access to code repositories. This level of privilege creates a "high-value target" scenario. If an adversary compromises an agent’s instructions, they could potentially inject malicious code directly into a production environment.
The implication is clear: identity and access management (IAM) must evolve to include "Agent Identity." Just as we manage human identities with roles and permissions, organizations must develop sophisticated, granular control policies for their digital agents.
The Path Forward
For organizations looking to scale their AI efforts, the guidance from experts like Ceylan points toward a balanced, risk-aware approach:
- Iterative Deployment: Do not deploy autonomous agents in mission-critical environments until they have been rigorously stress-tested in sandbox environments.
- Human-in-the-Loop Oversight: Regardless of how efficient an agent is, maintain a "break-glass" procedure where humans retain final veto power over critical system changes.
- Model Agnosticism: Avoid becoming tethered to one vendor’s ecosystem. A multi-model strategy is the best defense against model-specific vulnerabilities.
- Continuous Monitoring: Since agents operate in real-time, the security monitoring must also be real-time. This requires specialized tools that can track agent behavior and flag deviations from the "norm."
Conclusion
The era of the agentic enterprise is not coming—it is already here. With 83% of enterprises already running these systems, the competitive advantage will go to those who can master the art of the "secure-by-design" agent.
While the risks are non-trivial, the potential for agents to augment human intelligence and provide a robust defense against modern threats is equally significant. As Box’s experience demonstrates, the key is not to fear the autonomy of the machine, but to channel it through rigorous, multi-layered governance that ensures the human element remains the final arbiter of truth and security.
The road ahead is complex, and as Ceylan noted, it will require constant reassessment. Yet, for those who successfully navigate the intersection of speed and security, the payoff will be a highly efficient, automated enterprise capable of operating at the speed of thought.
About the Author
