A few principles of Agentic Security

 


  1. smarter agents are less likely to get duped.
    1. Though it is also true that an agent that doesn’t do its job well is also bound to be bad at doing an evil job well….
  2. shared experiences / many eyeballs
    1. This tends to align with the design principle to meet your user where they live, as that often means slack, for example.
  3. deadly triad
    1. Segregate agents by what they can read and write, based on whether you would categorize it as “Sensitive internal data”, “untrusted external content”, and “exfiltration capabilities”
  4. fourth: defense in depth –   imagine stacking slices of swiss cheese, which will reduce probabilities that a threat can permeate 
    1. This might include sub-agents that summarize web fetches/ sensitive data rather than showing it verbatim
    2. deterministic regex hooks 
    3. etcetera
  5. If some tool is too dangerous and too powerful, like bash usage for an agent, make it user approved each time. 
  6. fifth: security through obscurity 
    1. Windows had viruses and macs didnt, historically, because there were so many more windows then macs. This is similar to what this principle means
  7. sixth: reduce volume
    1. Dont have too many agents running autonomously  


No comments:

Post a Comment

The most surprisingly reasonable lawsuit ever [part1]

When I set out to write this post, it was going to be called "The dumbest lawsuit ever". After some The Netflix challenge was a ...