AI Agents: The New Exposure Point in Your Corporate Infrastructure

AI Agents: The New Exposure Point in Your Corporate Infrastructure

The Invisible Threat: When Attacks Bypass the Endpoint

For a long time, cybersecurity defense has been heavily focused on the endpoint. However, modern attack patterns increasingly circumvent this line of defense. According to N-able’s recent “State of the SOC” report, in 2025, 18% of alerts originated from network and perimeter infrastructure. Even more concerning is that around half of all attacks never even touched the endpoint. This statistic is alarming on its own: organizations relying exclusively on endpoint monitoring would have missed as many as 137,187 network and perimeter threats across a 12-month period.

In a corporate context, this means that traditional security strategies are no longer sufficient. AI agents, which often operate via cloud-based APIs or internal network services, fit perfectly into this blind spot. By their very nature, they exist at the network layer and within the application logic, allowing them to remain invisible to endpoint protection software.

Do you have a question about AI security? You can reach us here:

The AI Agent as the Weakest Link

Small and medium-sized businesses (SMBs) and managed service providers (MSPs) are likely to be enthusiastic adopters of AI technology, as these tools help them achieve more with fewer resources. AI can help automate repetitive processes, summarize information, and support customers with fewer manual steps. The challenge is that these same organizations may still be maturing the processes needed to track AI infrastructure, patch open-source components, or monitor agent activity.

In many cases, the AI agent can represent an easy route into the systems and data behind it. A real-world example of this is the “BadHost” flaw discovered in the Starlette framework. This vulnerability demonstrated how weaknesses in underlying frameworks can potentially expose private AI tools, MCP servers, large language models, and the sensitive data those tools can access. The agent, designed to increase efficiency, thus becomes the newest and least-protected element of the attack surface.

AIQ Analysis: What This Means for European Businesses

From an AIQ standpoint, the situation described above affects the European corporate sphere, including Hungary, on multiple levels, from regulatory compliance to practical cybersecurity.

OWASP LLM Top 10 and Practical Risks

The “BadHost” flaw is a perfect illustration of several critical points in the OWASP LLM Top 10. Primarily, it embodies the risk of LLM05: Supply Chain Vulnerabilities, as the flaw was not in the AI model itself but in an external, open-source component (Starlette). It is also closely related to the problem of LLM08: Excessive Agency: if an agent running on a vulnerable framework has overly broad permissions, the impact of its compromise can be catastrophic, potentially leading to the leakage of sensitive data (LLM06: Sensitive Information Disclosure).

EU AI Act and GDPR Compliance

The European Union’s AI Act imposes strict cybersecurity requirements on developers and operators of high-risk AI systems. Using an inadequately audited and unpatched software component (as in the Starlette case) can directly violate these regulations. From AIQ’s perspective, maintaining and regularly reviewing a Software Bill of Materials (SBOM) will be essential for demonstrating compliance.

From a GDPR perspective, a data breach occurring through a vulnerable AI agent can have severe consequences. If the agent has access to personal data and an attacker gains entry through this vulnerability, the data controller’s liability is unquestionable. The lack of proper access management and network segmentation could be considered an aggravating factor in an official investigation.

The Fundamentals of Defense: Old Principles in a New Context

The key takeaway is that defending against the new attack surfaces opened by AI tools does not require some mystical new technology, but rather the extension of well-established security fundamentals to these new systems. Based on AIQ’s audit experience, the following areas require special attention:

  • Asset Inventory: You must know exactly which AI models, agents, and supporting software components are running in your environment.
  • Vulnerability and Patch Management: The use of SBOMs and the continuous scanning and patching of open-source dependencies are critical.
  • Identity and Access Management (IAM): AI agents must also operate on the principle of least privilege. They should only have access to what is absolutely necessary to perform their tasks.
  • Network Segmentation: AI components should be placed in isolated network segments to prevent a potential compromise from spreading across the entire corporate network.
  • Monitoring and Logging: The activities, API calls, and data access of agents must be continuously monitored to detect anomalies.

AI security could also become a new advisory and managed service opportunity for MSPs. Those providers who can take on these complex security tasks for their clients can gain a significant competitive advantage in this rapidly evolving market.

Attila Rácz-Akácosi

Independent AI Security Specialist

Two decades of analytical and systems-oriented experience. I have been working with artificial intelligence since 2017. In recent years, I have specialized in AI/LLM security and AI Red Teaming. Systems-level thinking instead of endless vulnerability checklists.