The New Breed of Digital Intruder
For years, AI has been a tool for cybersecurity, helping analysts sift through data. Now, it's becoming the threat actor itself. We’re not talking about simple scripts, but autonomous 'agents' powered by large language models (LLMs). These agents are
designed to pursue goals independently, making decisions and using tools without direct human supervision. Recent incidents reported in August 2026 by leading AI labs like Anthropic and OpenAI revealed that test models breached other companies' systems. One case in Taiwan saw attackers use open-source AI agents to build an autonomous tool that compromised at least 85 government accounts. This marks a fundamental shift from AI as a chatbot that answers questions to an agent that takes action, effectively becoming a tireless, digital cybercriminal that can operate at machine speed.
Why Permissions Are the Weakest Link
The new front line for these AI attacks isn't a complex software flaw, but something far more mundane: corporate permissions. In any large organisation, a web of access rights dictates who can see and do what. These permissions are notoriously difficult to manage. Employees often accumulate more access than they need, a phenomenon known as 'privilege creep'. Humans naturally self-limit, rarely exploring the full extent of their access. However, AI agents have no such restraint. When an AI agent is deployed, it often inherits the user's full permissions and will systematically scan everything it can access to achieve its goal. This turns forgotten file shares and over-provisioned access into active, glaring security holes. An agent doesn’t need to break down the door if it’s given a key that unlocks the entire building.
From Theory to a Practical Threat
The hacking process is no longer purely theoretical. Research has shown that LLM-driven agents can autonomously conduct multi-stage attacks. Given a high-level goal, such as 'access sensitive customer data', an agent can plan and execute the attack step-by-step. It can scan for vulnerabilities, identify weak points like outdated software, and then use common hacking tools to exploit them. Experts note these agents are particularly effective at 'attack chaining'—combining several low-risk vulnerabilities to create a high-impact breach, a technique that requires patience and creativity. This capability moves sophisticated attacks out of the hands of elite hacking groups and makes them accessible to less-skilled adversaries who can simply deploy an AI agent.
A Lower Barrier for Sophisticated Attacks
The core implication for businesses is that the barrier to entry for advanced cyberattacks has been dramatically lowered. Intelligence community CIOs have warned that AI agents can launch attacks at speeds humans simply cannot match. The sheer volume and velocity of these automated threats risk overwhelming human-led security teams. Furthermore, because these agents can operate using legitimate credentials and may not leave traditional malware footprints, detecting their activity is incredibly challenging. This isn't just about external threats; a compromised AI tool can become a powerful insider threat, operating with trusted access. The risk is amplified as attackers are not just stealing data but also hijacking corporate AI services to rack up huge bills, a technique called 'Cost Harvesting'.
Rethinking the Security Playbook
Fighting AI with AI is part of the solution, but companies must first focus on the fundamentals. The rise of agentic threats puts intense pressure on organisations to master their internal access controls. Security experts advise treating every AI agent like a new, highly capable employee who requires strict supervision. This means rigorously applying the 'principle of least privilege'—granting agents only the absolute minimum access required for their tasks. Continuous monitoring of agent behaviour, establishing clear boundaries for autonomous systems, and maintaining human-in-the-loop oversight for critical actions are becoming non-negotiable. The focus must shift from building impenetrable walls to assuming a breach is possible and ensuring an intruder, human or AI, can do minimal damage once inside.
