This episode includes paid promotion. Security Unfiltered was paid by Eve Security for this sponsorship.
← All episodesYour AI Agents Will Try Every Permission You Gave Them | Nadav Cornberg Eve
Nadav Cornberg of Eve Security comes back for the later conversation in this Eve arc. We got into agent runtime security: understanding intent inside a session, session tainting when privileges should shrink, MCP connections that let tools take action in systems the author never planned for, and why static guardrails cannot keep up once agents start chaining actions on their own. I also talked through hooking an agent to my own GitHub repos and watching it ask for delete permission.
Key takeaways
- Agents will try every permission you gave them, and they will have a reason ready.
- Do not grant delete or other destructive scopes unless the job cannot be done without them.
- Session tainting means shrinking privileges in the same session when behavior goes sideways, for example after a second unexpected delete.
- A new MCP connection can reach whatever that tool can reach, including systems nobody planned to expose.
- Static guardrails are not enough once agents chain actions.
Transcript
Session tainting is the ability for you to look at a session and determine in that session if privileges should be expanded or reduced. I'll give you an example. I don't want bulk or destructive operations from happening. Let's not allow more than two deletes.
Nadav Cornberg, 24:06
More like this
Don't miss an episode
New conversations, in your inbox.