Define no-go boundaries
What agents must never touch
The Bad Tradeoff:
Overpermissioned or Useless
The Bad Tradeoff:
Overpermissioned
or Useless
If AI agents inherit broad employee access, they work well but
touch systems far beyond the task. That works in low-risk
environments, but not in sensitive systems.
If access is tightly
restricted, they lose the context needed to be useful.
If AI agents inherit broad employee access,
they work well but touch systems far beyond
the task. That works in low-risk environments,
but not in sensitive systems.
If access is tightly restricted, they lose the context needed
to be useful.
Kanuki defines agent permissions by what they must never be able to break: Kanuki definesagent permissionsby what they mustnever be ableto break:
Formal verification
Prove boundaries hold across the whole system, including indirect paths
Broad working context
Agents keep full working context, while critical systems remain protected
No permission babysitting
Agents operate within boundaries
without constant
adjustments
Decision traceability
Who delegated access, what
happened, and why
Examples of no-go boundaries
A. Critical actions:
Agents must never delete databases, initiate
payments,
or trigger deployments.
Agents must never delete databases,
initiate payments,
or trigger deployments.
B. Privilege escalation:
Agents must never grant themselves
admin rights.
C. Sensitive data:
Agents must never access consumer data.
How It Works:
Map the environment
and agent usage
Understand where agents run, what they touch,
and what
access they inherit.
Understand where agents run,
what they touch, and
what access they inherit.
Understand where agents run,
what they touch, and what access
they
inherit.
Define what agents must
never be able to break
Turn security requirements into explicit
no-go
boundaries.
Prove and enforce
those boundaries
Formally verify they cannot be crossed, then
enforce
access within them and log all actions.
Formally verify they cannot be crossed,
then enforce
access within them and log all actions.
Formally verify they cannot be crossed,
then enforce
access within them
and log all actions.
Experimenting with AI agents?
Let’s see what it would take to bring them
into production
on your terms.
Questions? Skeptical?
Talk to the founder:
Lisa Akselrod
Founder & CEO, Kanuki
