Agent SecurityRelated More in Secure
Tool Permissions & Least Privilege
Which agent may issue which call, at which scope, under which limit. Most agent deployments answer that question with 'all of them, unlimited'.
Sound familiar?
- One API token grants the agent everything the application can do.
- Write access was granted for one use case and never narrowed afterwards.
- No rate, cost or blast-radius limits on tool calls.
- Nobody has modelled what a compromised agent could reach in one hop.
What we do
Permission matrix
Agent × tool × scope × limit, written down and enforced rather than assumed.
Scope reduction
Read-only by default, write access only where a use case demands it.
Limits & circuit breakers
Rate limits, spend caps and automatic stops when behaviour goes out of band.
Blast-radius test
We simulate a compromised agent and show exactly how far it gets.
Questions we get
Won't least privilege break our agents?
It breaks the ones that were silently doing more than intended. We reduce scope in stages with monitoring, not in one cut.
Does this work with MCP tools?
Yes — MCP tool scopes are exactly where this matters most.
Tell us what's running in production.
We'll tell you what we'd check first — and what we wouldn't bother with.
Book a call