Practical AI Dealing with increasingly complicated agents
182 snips
Oct 16, 2025 Donato Capitella, a Principal Security Consultant at ReverseSec, shares his insights into the evolving landscape of AI and security. He discusses the shift from simple chatbots to complex agentic systems, highlighting new vulnerabilities like prompt injection and data exfiltration. Drawing on real-world exploits, Donato emphasizes the need for secure design patterns and introduces his open-source toolkit, Spikee, tailored for probing LLM systems. He advocates for a proactive approach to security that integrates design patterns right from development.
AI Snips
Chapters
Transcript
Episode notes
Agentic Workflows Are Becoming The Norm
- Agentic systems are now common: LLMs calling external tools form loops and orchestrators.
- Donato warns this shift turns many apps into agentic workflows with more attack surface.
Tools Become Public Attack Surfaces
- Any tool exposed to an LLM is effectively exposed to anyone who can control LLM input.
- Donato stresses that internal APIs turned into tools create new internet-facing attack surfaces.
Enforce Deterministic Authorization For Tools
- Enforce deterministic, non-LLM authorization for tool calls and identity passing.
- Do not rely on the LLM to decide whether a function can be called in a given context.