Autonomous AI agents in cybercrime

How does the use of autonomous AI agents change the threat potential in cyberspace from the perspective of protecting a company?

Sub-questions examined

1. 01Which realistic attack scenarios involving autonomous AI agents can be identified today?
2. 02Which organisational and technical weaknesses become particularly exploitable through such scenarios?
3. 03To what extent do human factors – convenience, lack of awareness, uncritical use – contribute to that exposure?
4. 04Which concrete safeguards and organisational changes are needed to make companies more resilient against AI-assisted attacks?

• OSINT exposure and missing visibility

  Companies often do not know what is visible about them from outside. The sources usually originate outside IT – in HR, marketing, communications – which is why ownership cannot be left with IT alone.

• Blurred processes and diffused responsibility

  Once exceptions become the norm, an attacker can anticipate that norm. Four-eyes approval, fixed escalation paths and exceptions recorded in writing only work if leadership does not quietly undermine them through pressure for speed.

• Detection gaps from missing correlation

  Individual events look harmless; the pattern does not. Without correlation across sources, slow and distributed attacks stay invisible.

• The human factor as an amplifier

  Convenience, time pressure and routine lower the threshold for bypassing safeguards – in every phase, not just in social engineering. The very relief that makes digital work viable becomes an attack vector.

The next step plausibly lies in multi-agent systems: instead of one generalist, specialised agents work together – reconnaissance, communication, technical paths – coordinated by something that prioritises goals. On the defensive side that suggests a picture of agent against agent. The research needed is less about new model tricks than about empirical studies of real incidents and protection concepts that also work on a small budget.

• The scenarios rest on technically plausible conditions, not on empirical validation through interviews, case studies or incident data of my own. The analysis is qualitative rather than statistical.
• The field moves fast. Some assessments may look different in one or two years – in both directions, because defences adapt too.
• Legal and economic depth were deliberately left out; each would deserve a thesis of its own.
• The work draws partly on vendor threat reports, which can be interest-driven. They were framed with research and official sources, but some bias may remain.