On-demand webinar: How an attacker views your AI security

Look past the model

Most AI security programs start with the model. Attackers start with what surrounds it: the tools, data, identities, and integrations your AI systems can reach.

With current large language models (LLMs), prompt injection and jailbreaking can’t be fully prevented. The question for security leaders isn’t whether the model is secure. It’s how much harm a manipulated model can cause inside your environment.

Reversec’s Donato Capitella and Nick Jones discuss AI security from the attacker’s seat. They share examples from client work and set out a practical way to decide what to secure first.

Key questions answered:

• Why does testing the model rarely reduce real risk, and where should you focus instead?
• How do attackers exploit AI integrations, from chatbots with database access to back-end agents that process untrusted files?
• Has AI changed how organizations get compromised, or made existing attacks cheaper and faster?
• Why can AI agents behave like insider threats?
• Which controls limit the blast radius when an agent is manipulated?
• How do you prioritize when every team is adopting AI at once?

Speakers

Donato Capitella, is a Principal Security Consultant at Reversec, specializing in the security of generative AI, large language model applications, and AI agents. His work focuses on prompt injection, guardrail testing, and secure agent design. He is also the creator of Spikee, an open-source toolkit for evaluating prompt-injection vulnerabilities, and regularly shares his research at international security conferences.

Nick Jones, is the Head of Research at Reversec. He led Reversec’s approach to cloud security in large enterprises and cloud-native organisations, and now focuses on research into the security of AI-powered systems. A 12-year cybersecurity veteran based in Copenhagen, Nick has previously spoken at numerous conferences and events including fwd:cloudsec US, DEF CON, Disobey, V2 and Blue Team Con.