Read here,
write there.
Two approved tools with different scopes create a bridge between them. The classic pair is read-confidential plus send-external; there are usually several less obvious ones.
MKC2 results ↗Privilege escalation in agent systems rarely comes from a single dangerous permission. It comes from a combination that no individual review considered.
01 / DEFINITION
The short answer first.Privilege escalation in an AI agent system is any path by which the agent achieves an effect beyond what was approved. In practice it usually arises from composition — several individually reasonable capabilities combining — rather than from a single over-broad grant.
The reason this differs from conventional escalation is that the agent composes its own paths. No exploit is required: the model is doing exactly what it was built to do, which is to find a sequence of available operations that accomplishes an objective. If the sequence happens to cross a boundary the reviewers never considered as a sequence, the system has escalated without anything malfunctioning.
WORKED EXAMPLE
A helpdesk agent can read employee account status and request an access review. A separate identity service can actually grant a role. Trouble arises if the service treats any request from the agent's application account as an authorised grant. A user with an ordinary support ticket may then cause a privileged action through a more trusted intermediary.
Create a test employee who may request access but cannot approve it. Let the agent prepare a request for a role the employee does not hold. Verify that the identity service receives the original identity, requested role, target account and approval status. A service-to-service connection proves which application called; it does not prove who approved the privilege.
Run a legitimate control with an authorised administrator approving that exact role for that exact employee. Then reuse the approval with a different role, target account or expiry time. Each altered operation should require its own valid authorisation. Capture the final identity-system state rather than accepting a chat response that says the request was refused.
Check indirect execution as well. A queue worker, automation rule or retry service might run with broader privileges than the original agent. Revoke the approval while the request is queued and confirm that the executor evaluates current authority before applying the change. If the employee gains the role after revocation, identify the component that trusted stale approval.
Give operators a safe way to investigate the attempted escalation without publishing employee details in an unrestricted log. The useful record links the request, decision, executor identity and final role assignment. It also tells the team whether the failure came from model planning, an overly broad service grant or an executor that failed to check the underlying user's authority.
02 / FOUR PATTERNS
Two approved tools with different scopes create a bridge between them. The classic pair is read-confidential plus send-external; there are usually several less obvious ones.
An agent with broader rights than its caller performs an operation on the caller's behalf, and the caller's own restrictions never apply.
A sub-agent is granted marginally more than it strictly needs, and a chain of such grants ends far from the original scope.
A fourth pattern is persistence. An agent with memory, a scratchpad or a writable store can carry information or instructions from one session into another, which crosses the boundary between a run scoped to one user and a later run scoped to a different one. Memory that survives a session should be treated as a channel with its own access controls, not as an implementation detail.
03 / REVIEWING THE SET
Translate every capability into what it lets the system do to the outside world: read this class of data, change this record, reach this destination.
For each pair, ask what becomes possible that neither allows alone. Most escalation paths appear at this step and nowhere else.
Whose rights apply at each call, including inside sub-agents and inside tools that call other services on the agent's behalf.
Adding one tool changes every pair. A capability set approved last quarter is not approved now.
Step two is tedious and is where the value is. A set of eight tools has twenty-eight pairs, which is a reviewable number, and the interesting findings are usually concentrated in three or four of them. Automating the enumeration — generating the pair list and asking the owner to justify each combination — makes it a routine rather than an exercise of memory.
04 / STRUCTURAL PREVENTION
Some architectures remove the escalation path rather than reviewing it.
The first pattern is the most generally useful and the least adopted, because it requires designing the interface between components rather than handing one model every tool. The cost is a schema and some engineering; the benefit is that the escalation surface stops growing with each new capability.
05 / QUESTIONS
No vulnerability is involved. The agent uses approved capabilities in an unapproved sequence, which it discovers itself while pursuing a goal. Nothing malfunctions, so conventional detection based on failures and anomalous authentication sees nothing.
An agent with broader rights than its caller acts on the caller's behalf, so the caller effectively borrows the agent's privilege. Propagating the requesting user's identity through every call removes it. See AI agent permissions.
Yes, in the disclosure direction. Read plus any outbound channel is an exfiltration capability, and rendering a URL can be enough. Read-only bounds the write risk and does not bound the disclosure risk.
Materially, because delegation chains multiply the number of identity transitions and each one is an opportunity to widen scope. Requiring that delegation can only narrow, and recording the full chain, keeps it tractable.
On every change, because one new tool changes every pair. A scheduled quarterly review of a set that changes weekly is reviewing something that no longer exists.
WORK WITH PRIVILEGE AI
Review what tools can do together.
We look at how an agent's tools combine, because two harmless tools can add up to an unsafe one.
Firm limits instead of judgment calls.
We build agents whose reach is fixed by the design rather than left to the model's discretion.
Fewer surprises in production.
Agents that are boundaried by design are easier to approve and safer to run.
Talk with Privilege AI
Describe the low-privilege task and the higher-privilege service you want to check for unintended delegation.
CONTACT PRIVILEGE AI
Describe the low-privilege task and the higher-privilege service you want to check for unintended delegation.
Privilege AI designs agent boundaries where the escalation surface is a schema rather than a model's discretion.