
AI-powered cyberattacks are no longer only a laboratory possibility. In November 2026, the European Union confirmed that a campaign supported by artificial intelligence tools affected one of its institutions and led authorities to take several applications temporarily offline. The incident offers a practical warning for any company connecting agents to data, APIs, and internal systems.
The available information requires precision. European officials said the investigation was ongoing and that they had not detected a data breach affecting staff as a whole. The actor behind the incident has not been identified publicly. What was confirmed was the use of AI tools to help detect and exploit vulnerabilities—not an unlimited machine acting without people or constraints.
What changes with AI-powered cyberattacks?
Artificial intelligence does not replace traditional methods. Stolen credentials, weak configurations, unpatched software, and excessive permissions remain common entry points. What changes is the speed at which an attacker can collect information, analyze a surface, adapt tests, and repeat attempts.
A recent report from the U.S. Government Accountability Office warns that AI can amplify the reach and impact of known attacks. It also identifies agentic AI executing several stages of an attack with minimal human input as an emerging trend.
The conclusion is not “disconnect every AI system.” It is more specific: every connection given to an agent should be treated as an operational capability requiring identity, permissions, limits, and evidence.
When a convenient connection concentrates too much power
Imagine an internal agent that checks inventory, prepares quotations, and updates sales opportunities. To save time, the team gives it one credential with broad database and CRM access. The demonstration works, but nobody documents which tables it may read, which records it may modify, or what happens if it receives a manipulated instruction.
The problem is not only the model. It is an architecture that cannot answer basic questions: who performed the action, which credential was used, what changed, can it be stopped, and can the change be reversed?
1. Give each agent its own identity
Do not connect an agent through a personal administrator account. Each agent should operate with a distinct identity, dedicated credentials, and a recognizable function. This allows access to be revoked without blocking an entire team and makes each action attributable to its real origin.
2. Apply least privilege from the beginning
A read-only agent does not need permission to delete records. A scheduling agent should not access financial information. Divide capabilities by task and avoid granting permissions “in case they are needed later.” Authorization should expand only when a verified need exists.
3. Separate reading, proposing, and executing
Not every action requires the same degree of autonomy. A safer architecture distinguishes three levels:
- Read: retrieve the information required for a task.
- Propose: prepare a change for review.
- Execute: modify a system or send information.
Payments, discounts, deletions, permission changes, and sensitive communications should require approval or additional rules.
4. Protect MCP, APIs, and tools as access points
MCP can make it easier for an agent to discover and use tools, but it does not replace authentication or authorization. Every server, API, or integration must validate who is requesting an action, what operation is being attempted, and which resources are involved.
Keep secrets out of prompts, histories, and shared files. Rotate credentials, limit their lifetime when possible, and store them in systems designed for secrets management.
5. Record decisions and actions, not only errors
A system can finish a task without raising an exception and still produce the wrong outcome. Useful records include the request, the selected tool, applied permission, executed action, result, and identity that approved an exception.
This traceability supports incident investigation, quality measurement, and decision reconstruction without relying on a person's memory.
6. Define limits and an immediate stop mechanism
Set limits for consumption, frequency, data volume, and number of actions. An agent that normally processes twenty requests per hour should not execute thousands without triggering an alert.
It also needs an independent stop mechanism: revoke credentials, pause an integration, or switch the agent to read-only mode.
7. Test malicious instructions and real failures
Do not test only the expected path. Include messages attempting to override rules, documents containing hidden instructions, incomplete data, unavailable APIs, duplicate responses, and unauthorized users. Verify that the agent refuses, escalates, or stops as designed.
Checklist before connecting an agent
- Does it have a distinct identity and dedicated credentials?
- Can it access only the data required for its task?
- Do critical actions require approval?
- Are secrets kept out of prompts and logs?
- Does every action leave reviewable evidence?
- Are there volume, cost, and frequency limits?
- Can access be stopped and revoked within minutes?
- Have manipulated instructions and integration failures been tested?
How Ainoa can help
Ainoa builds Custom Intelligent Agents around specific processes, with integrations, permissions, approvals, and escalation paths defined for the real operation.
Custom Administrators can centralize statuses, histories, metrics, exceptions, and controls. The goal is not to promise zero risk. It is to prevent automation from becoming a black box without owners or limits.
Frequently asked questions
Can AI conduct a cyberattack without human intervention?
Some systems can automate several stages, but incidents must be assessed individually. Automation does not prove that the entire operation happened without human planning, tools, or decisions.
Is connecting an agent through MCP unsafe?
Not necessarily. Risk depends on authentication, permissions, exposed tools, input validation, logging, and the ability to revoke access.
Do small companies need these controls?
Yes, adapted to their scale. Separate accounts, limited permissions, approval for critical actions, and useful logs are not exclusive to large organizations.
The advantage will come from better control
AI-powered cyberattacks increase pressure on organizations that still manage access and integrations informally. The response is not to halt every automation initiative, but to design controls from the start.
If you want to connect agents to real systems without losing visibility over permissions and actions, talk to Ainoa. We can help turn a process into measurable, traceable automation with human intervention where it matters most.