The recent developments surrounding vulnerabilities in major AI repositories like Hugging Face serve as a critical wake-up call for the cybersecurity community. As we accelerate toward an agentic future, the platforms we rely on for innovation are increasingly becoming the primary vectors for systemic risk.

From Shadow IT to Shadow AI

For years, CISOs fought the battle of Shadow IT. Today, we face a far more insidious challenger: Shadow AI. This is the unsanctioned use of public AI tools and unvetted open-source models by well-meaning teams who, in their drive for efficiency, inadvertently compromise sensitive organizational data and trust.

The incident involving Hugging Face highlights the fragility of our AI supply chain. When we pull a model from a repository, we aren't just importing code; we are importing a "black box" of logic that can be exploited via prompt injection, data poisoning or even integrated into fully autonomous ransomware attacks.

The Rise of Agentic Risk

We are moving beyond simple chatbots to AI agents, autonomous systems capable of executing transactions and making decisions. While this agentic edge promises resilience and innovation, it also creates a massive governance gap.

Recent research indicates that less than half of businesses currently have an AI governance policy in place. This lack of oversight is what allows a vulnerability in a single model repository to ripple across an entire enterprise.

In addition, those enterprises that have moved AI agents into production are concerned about how these AI agents behave, with only 34% of technology decision-makers at director level and above globally only trusting the actions their AI agents are taking.

Building a "Human-in-the-Loop" Roadmap

Securing this new frontier requires more than just technical patches; it requires a fundamental shift in how we manage human risk.

• Human-Centered Vetting: Establish human-led assessment protocols that require expert review of all AI models and vendors before deployment, ensuring every integration is intentionally approved by human oversight.
• Continuous Human Oversight: Implement "human-in-the-loop" controls that mandate human intervention for critical decision points, ensuring that autonomous agents remain aligned with organizational parameters and never operate without supervision.
• Empowered Workforce Security: Evolve security training to prepare employees not just to spot threats, but to serve as active monitors of the AI agents they manage, shifting from passive users to empowered human safety mechanisms.

The Path Forward

The goal is not to stifle innovation but to empower it with integrity. We need to move from a state of "uncontrolled AI" to a practical roadmap of risk assessment and secure implementation.

In my recent talks, I provide a quick framework for the audience called EEG: Embrace, Educate and Govern.

• Embrace: Don't stifle innovation; facilitate it. Integrate tools that allow you to inventory and assess the AI agents currently operating within your ecosystem.
• Educate: The focus must shift from merely training humans to spot phishing to managing the risks posed by the digital agents those humans interact with every day. Organizations should leverage AI-native security awareness training that adapts to the real-world threats facing your users.
• Govern: Vetting is non-negotiable. Organizations must establish rigorous vendor and model vetting processes before any "off-the-shelf" AI is integrated into workflows. By utilizing Agent Risk Manager, you can gain the visibility needed to understand how AI agents behave in real-time, ensuring that "human-in-the-loop" guardrails are always in place.

As we saw at the recent Workforce Security Summit, the tools to manage these agentic risks are evolving. The question is no longer if your team is using these tools, but how you are managing the risk they bring to your doorstep.

By unifying your email security, security awareness training, and AI agent defense strategies, you can ensure your workforce remains an asset, not a vulnerability.