The aggressive adoption of AI will unlock incredible productivity and efficiency. However, recent news about Anthropic's new Claude model (which has successfully “escaped” from its sandboxed test environment) proves that agent powers are extremely difficult to manage, predict, and understand. As a result, organizations need to think more holistically about the appropriate controls, oversight mechanisms and governance that both companies and boards should have in place.
We must recognize so-called “human error” when a sophisticated, skilled, leading model company like Anthropic proves a risk to us. Of course, we know this from our experience with cybersecurity and the endless phishing emails we receive every day. Fact is that 94% of all breaches are caused by human error, such as clicking on an infected link. Just as we must continually train ourselves and our employees to avoid cyber-attacks, we must contend with force multipliers as we improve the skills of our workforce. Statistics show that 55% of all jobs in the United States will be reshaped, redefined, and reimagined by incorporating AI and agent capabilities into workflows.
This means we are all on the same learning curve. We encourage our employees to master AI and agents. As individuals, we use AI assistants in our daily lives. Many of us are now beginning to boldly use “vibe programming” using natural language prompts to create our own agents. This significantly increases your risk, as your attack surface has just expanded unless you proactively implement the necessary guardrails and protections.
Corporate boards need to think about acceptable use cases and absolute “no-go zones” for AI. For example, there should be no unsupervised AI decisions in highly regulated or safety-related fields. There also needs to be a clear structure that defines where the oversight of the AI will reside. Typically, this may fall to the audit committee, where cybersecurity is often the responsibility. However, given the breadth of audit committee powers and significant workload, it may be best to consider having a technology committee responsible for AI and cyber oversight, depending on how important AI and cyber oversight are to a particular company.
Implement comprehensive AI governance
For effective AI governance, boards need to identify an acceptable layer of risk. When training employees to be AI proficient and agent-ready, you need to set clear boundaries on how far AI can go.
for example, low risk The use case will be “read-only”. AI cannot create or publish content externally.
in medium risk In this scenario, we set certain limits on what is automated and ensure that the AI is structured in such a way that any mistakes or actions can be undone.
for very high risk In areas where actions can have irreversible and significant effects, such as intellectual property, payments, and data deletion capabilities, you should insist on “human participation” approval requirements.
Once the risks and use cases have been categorized, the individual owners of each need to be identified. Every agent must have an owner who is responsible for maintaining the agent's objectives and appropriate guardrails. Board members must request a “Representative Registration” that identifies the owner, the risks involved, risk tiers, and a description of the agent's capabilities.
A common mistake when it comes to guardrails is relying on paper policies. Policies are useless without physical evidence and active monitoring. Technical controls and testing are required to prove that guardrails are monitored and enforced.
This is a perfect metaphor for how cybersecurity monitoring works. Just as there is continuous monitoring in cyber, agent behavior must also be continuously monitored. A formal response and incident alert process is required, including an escalation path and, most importantly, a way to contain the damage and disable the agent. The same principles apply when considering attack surfaces and mitigations.
- Identity management: Every agent must have their own unique credentials. API keys cannot be shared.
- mapping: Each agent must be mapped to a human owner.
- Least privilege: Design everything with the least amount of privileges needed. Explicitly identify the systems, data, and endpoints that agents can access and the operations they are allowed to perform.
- Vulnerability points: Make sure you don't have direct access to the production database or admin API.
Essentially, your mindset should be either zero trust.
AI is a big opportunity, but also a big risk. The board must oversee this using the same considerations used for capital allocation, major acquisitions, and other high-stakes processes. Using a cybersecurity framework can help categorize the components needed for monitoring, such as data, privacy, ethics, compliance, and ensuring that AI algorithms are not “drifting.”
These are all important considerations for boards as they seek to develop a practical approach and checklist for operating proactively this year. With 55% of jobs expected to be AI-enabled, employees will quickly move up the learning curve to create and deploy agents for internal and external use. This makes AI governance an urgent priority to ensure enterprise security while remaining competitive.
