This is not science fiction, it is the situation today: publicly available AI models are in the hands of attackers. The same tools used productively also accelerate attacks across the entire attack chain.
“The new generation of AI models is able to identify vulnerabilities automatically and exploit them at the push of a button.”
Claudia Plattner, President of the BSI, BSI-Magazin 2026/01
“AI systems can now take on tasks that previously required experienced experts.”
Fred Heiding, security researcher, Harvard University
The threat landscape
AI accelerates every stage of the attack chain:
Target profiles from public sources, automated and complete.
Models analyze foreign code and find even long-overlooked gaps.
From CVE to a working exploit at the push of a button, no specialist team.
Polymorphic on demand: every delivery a one-off.
Flawless, personalized phishing campaigns in any language.
Voice, image, and video, convincingly real, live in meetings.
Multi-stage attacks, semi-autonomous and scaled to hundreds of targets in parallel.
AI assistants and agents become an attack surface themselves, e.g. via prompt injection. Test and harden AI systems.
More attackers, more attacks per attacker
Professionals scale with semi-autonomous agents: one operator no longer handles a single target but hundreds in parallel. At the same time the barrier to entry drops, whoever copied scripts yesterday now runs campaigns at a previously professional level. The attacks are not getting more sophisticated. There are simply more actors hitting more targets faster.
Risk is probability times impact. As the effort and cost per attack fall, the likelihood rises, and with it the risk. The gap no one exploited for three years gets found now. The homework you postponed is due now.
AI on defense: necessary, but not sufficient
Defensive AI is useful: pattern recognition, prioritization, and faster response. Against fully automated attacks, however, it stays symmetrical, AI versus AI, a race with no lead.
The difference is the human. The serious adversary is not fully automated: a human leads, the AI scales. Creativity and context beat any pure automation, and the balance tips toward the attacker.
Against a human who wields AI, no off-the-shelf product helps, only a prepared organization.
Where to start right away
No “AI-powered” miracle tool, but processes with clear ownership. Six steps that have proven themselves:
1. Assess where you stand and get the basics right
Know where you stand first, then act deliberately: shorten the patch window, MFA everywhere, hardening, tested backups. The free CyberRisikoCheck gives you the starting point, the assessment developed by the BSI per DIN SPEC 27076. Two hours of effort, an evaluated report with prioritized measures, no obligation.
2. Reduce your attack surface
Inventory, exposure, shadow IT. What you don’t know, you can’t patch.
3. Build and live an ISMS
Steer risks systematically instead of firefighting. Under NIS-2, mandatory for many anyway. Building an ISMS.
4. Practice for the incident
A business continuity management system, rehearsed before it burns.
5. Govern your own use of AI
Governance, least privilege, logging, and human approvals.
6. Have effectiveness tested
Regular penetration tests and attack simulations, AI-assisted like the attacker. For your own AI applications: AI penetration tests.
The attackers have gotten faster. So should you. We help you align your security with the new situation and prove that it works.