The prospect of what artificial intelligence could become has shifted our attention away from the immediate security risks facing businesses. We hear warnings about super intelligent systems escaping human control, AI models developing minds of their own, and autonomous agents making decisions nobody can predict. These possibilities make for compelling debate, but amid the inflated hype, businesses risk overlooking threats that are far less dramatic and much closer to home.
An AI agent with access to the wrong database, permission to execute an unintended command, or the ability to move between sensitive systems can cause significant damage without ever developing a mind of its own.
As businesses increasingly deploy agents across their operations, these are the risks that deserve attention. And while AI brings new complexities, securing it doesn’t require us to reinvent cybersecurity. Many of the answers lie in good engineering principles we already understand.
An AI agent is still a tool
An agent is something that takes actions to achieve set goals. In that sense, the concept isn’t particularly new. People are agents too: they make decisions, interact with systems and carry out tasks on behalf of organisations.
If you are the CEO of a company employing thousands of people. You wouldn’t hand every employee a badge that opens every door or give everyone unrestricted access to every database. Someone in marketing doesn’t automatically need sensitive financial records, and a graduate engineer shouldn’t have unrestricted control over a production database. The same logic applies to AI agents.
An agent answering customer queries needs relevant information, but not necessarily permission to modify customer records or access financial systems. Its permissions should reflect the job it’s been assigned, rather than everything the technology is capable of doing.
This is the principle of least privilege, a cornerstone of cybersecurity for decades. It’s particularly important with AI because agents can interact with multiple tools and systems, sometimes in unexpected ways.
Giving an agent an instruction doesn’t guarantee it will execute that instruction exactly as intended. But if we’ve restricted its access appropriately, an unexpected action is far less likely to become a serious security incident.
Engineer for the things that could go wrong
In traditional engineering, we don’t assume systems will never fail. We identify potential failures, assess their consequences, and put safeguards in place. AI agents deserve the same treatment.
Take an agent operating inside a company’s network. While completing a legitimate task, it attempts to reach a server outside its designated environment. Perhaps it tries to retrieve confidential information or execute a command it wasn’t authorised to run. Whether the agent has any concept of malicious intent is irrelevant. What matters is what it can do and the damage that could follow.
A conventional risk assessment provides a starting point: identify what could go wrong, assess the likelihood and potential impact, then put appropriate safeguards in place. For example, an agent could operate in an isolated environment, with network restrictions preventing it from accessing sensitive servers. Actions carrying greater risks could require human approval. That way, even if an agent behaves unexpectedly or attempts something beyond its intended role, the damage it can cause is contained.
Businesses must also assess the remaining risk and revisit their controls as agents gain new capabilities. Connecting an agent to another database may seem like a small change, but it can significantly expand what it can affect.
Red-team exercises offer another familiar safeguard. By deliberately testing an agent’s boundaries, security teams can identify weaknesses before they cause problems in production.
This is classical risk management, applied to a different kind of system.
Not every problem needs an AI agent
Good engineering also means knowing when not to use AI.
Consider a company transferring information between two internal systems according to fixed rules. A conventional integration can handle that reliably, with clearly defined inputs, outputs, and permissions.
Replace it with an AI agent that interprets instructions, selects tools, and decides how to complete the transfer, and you have potentially introduced greater uncertainty, broader access requirements, and additional monitoring needs. If the outcome is no better, that complexity serves little purpose.
AI agents are valuable when tasks require flexibility, interpretation, or decisions that rigid workflows struggle to accommodate. But adopting them simply because the technology is available can create unnecessary security exposure.
Shadow AI reveals a different security challenge
Technical controls only address part of the problem. Organisations also need to understand how employees are using AI at work.
Look at shadow AI – the use of unapproved applications in the workplace. It is easy to label this as a compliance issue, but it is actually a productivity issue. Employees, more often than not, are not trying to break the rules; they are just trying to do their jobs. Whether it’s a developer writing code faster or a team member analysing information easier, people will always find a workaround if approved tools aren’t provided.
The risks are real, particularly when confidential information enters external services. But restrictions alone won’t address why employees are seeking these tools.
Businesses need visibility into which applications are being used and what information is moving in and out of their environments. They should also consider whether popular tools can be approved securely or replaced with suitable internal alternatives.
Education is equally important. Someone handling sensitive healthcare records faces different risks from an employee drafting general marketing content. Policies should reflect those differences, and employees need to understand the reasoning behind them.
Keeping pace with AI, grounded in the fundamentals
As AI agents become more capable, the challenge for businesses will shift from securing individual tools to managing entire networks of agents working across different systems, making decisions and taking actions on our behalf.
While the technology is evolving rapidly, the fundamentals of cybersecurity remain the same. Carefully defined permissions, rigorous testing, appropriate monitoring, and human oversight will be just as important in managing these increasingly complex systems. The difference is that businesses will need to think about security from the outset, rather than bringing engineering and security teams in once an agent is ready for deployment.
Developing the next generation of engineers will be equally important. Junior engineers, already accustomed to working with AI, need opportunities to work alongside experienced colleagues, combining their fluency with new tools with the critical thinking, technical understanding and security judgement that come from experience.
As businesses race to give AI agents more autonomy, they need to be equally ambitious about developing the people and practices that will keep them secure. After all, there’s little value in building an agent capable of doing extraordinary things if you can’t confidently control what it does.
That’s not science fiction. It’s good engineering.
Source: www.cybersecurity-insiders.com
