Penetration testing with frontier models: when security checks become a threat
The Guardian works for the attacker: how AI security tools are hijacked - and what you can do about it.
Topic
When & Where
Details
Format:
Lecture
Language:
German
Session description
SAST scanners, SOC Copilot, Triage Agent: You’re buying AI tools designed to detect attacks - and ending up with a new point of entry. That’s because these tools inherit every known LLM vulnerability. They’re manipulated, of all things, through what they’re supposed to analyze:
logs, tickets, code, PDFs. A single crafted line is enough to make the guardian work for the attacker - classic AppSec controls are ineffective here. On top of that, there’s a dangerous gap: Who actually checks the scanner, the copilot, the agent? A gap in responsibility arises between the vendor, the CISO, the DPO, and the AI Officer - and that’s exactly where attackers strike.
The presentation uses reproducible attack chains to demonstrate how this works - and provides a practical testing and ...
