Small Agents, Smaller Blast Radius (isc.sans.edu)

In a thought-provoking critique, Mark Baggett argues that the prevailing approach in AI security—encasing powerful AI agents in protective layers—is misguided. Instead of attempting to control these general-purpose AI tools by limiting their capabilities through containers and policies, Baggett advocates for creating smaller, single-purpose agents with strict access to only what they need to function. This aligns with the principle of least privilege and emphasizes that if even the developers struggle to manage their AI's behaviors, the industry should reconsider how much power these agents are granted. Baggett highlights the potential pitfalls of existing security measures, such as Check Point's Dangerous Deviation Detection, which evaluates actions proposed by AI agents before execution. While such a detector can identify dangerous actions, Baggett questions the fundamental design of assigning broad permissions to investigation agents. By separating duties and restricting access, organizations could significantly reduce risks associated with AI misinterpretations and overreach. His vision encourages a more responsible development of AI in cybersecurity, where powerful coding agents can be employed to create focused agents, thus improving security outcomes while limiting potential damage, ultimately transforming how AI capabilities are managed and deployed.