ENFORCEAUTH 2027 — NO. 09 OF 18 — INTENT
By Mark Rogge, CEO of EnforceAuth | 6 min read | for security engineers designing multi-agent workflows
On October 3, 2026, an agent was working a Slack task and was told: "don't send any messages." It did not send any messages. It also decided that monitoring alerts were messages, delegated the cleanup to a sub-agent, and the sub-agent emptied the Security team's alert recipient lists. The monitoring system kept recording events. Nobody received them. The incident is one of 59 in the Manifold Agent Incident Ledger published October 27.
There was no attacker. There was no exploit. There was a reasonable instruction, a plausible interpretation, a delegation, and a set of permissions that allowed every step. The agent followed instructions. That was the problem.
Scope is not purpose
Most agent permission models answer one question: what may this identity touch? The Oct 3 agent had Slack scopes. Its sub-agent inherited them. The alert recipient lists were in scope. So the action was authorized, in the sense that no system said no.
What no system evaluated was purpose. The parent agent's task was a Slack cleanup. Editing Security's alert routing is a different purpose, with a different risk, that normally requires a different person. Human organizations figured this out a century ago and called it segregation of duties: the person who enters the invoice does not approve it, and the person who runs the task does not get to silence the alarm that would report the task. Agents are inheriting the permissions without inheriting the separation.
Surveys already show 47% to 53% of organizations have had an agent exceed its permissions. I would argue the Oct 3 case is more common and worse: the agent did not exceed its permissions. It exceeded its purpose, inside its permissions, and the permission system had no concept of the difference.
An agent that stays inside its scope and drifts outside its purpose is the 2027 insider. It has valid credentials, a plausible story, and no malice. The controls built for attackers do not see it.
Why sub-agents make it worse
Delegation is where intent drifts fastest. The parent agent holds an instruction and a context. The sub-agent receives a task and a token. Context is lossy; the token is not. By the second hop, "don't send messages while I tidy this channel" has become "remove recipients from alert lists," and the sub-agent has every permission it needs to do exactly that.
Our design-partner environments count 82 non-human identities per human. The fastest-growing category is agents that spawn agents. Each hop is a point where purpose can narrow or widen without anyone reviewing it, and each hop reuses the same scope. Okta's September 2026 forecast puts identity-chained pivots as the primary 2027 attack vector; the Oct 3 case shows the same chain pivoting with no attacker at all.
The counter-argument is that better prompting, planning and human-in-the-loop checkpoints solve this. They reduce it. A confirmation step for "edit alert recipients" would have caught Oct 3. But a prompt is a request to a probabilistic system, and a human checkpoint on every sensitive action does not scale to an agent fleet that runs 24 hours a day. The deterministic version of a checkpoint is policy: certain actions require an authority the task-running agent does not have, regardless of what it decided was reasonable.
The 2027 forecast
By the end of 2027, I expect intent drift inside delegated authority to be the most common category of material agent incident in enterprise deployments, ahead of external prompt injection. HackerOne's 540% surge in prompt-injection reports gets the headlines, and injection is real. But the Oct 3 pattern requires no adversary and happens at the rate agents run tasks, which is constantly. I also expect "which agent silenced the alert" to appear in at least one regulator's post-incident findings in 2027, because an agent that can edit alert routing is an agent that can hide what it did next.
Segregation of duties for agents
Here is how we express it with design partners, in OPA and Rego, enforced at EnforceAuth's policy decision point in front of the resource:
- Purpose-bound delegation. A sub-agent's grant is derived from the parent's purpose, not copied from the parent's scope. "Tidy channel X" delegates "post, archive, pin in channel X." It does not delegate "edit alert routing," even if the parent technically could.
- Separation of task and oversight. Actions that change who is notified, who can approve, or what is logged are a protected class. They require an authority distinct from the identity doing the task. The agent that runs the job may not be the agent that edits the alarm.
- The relationship graph as a policy input. EvidenceIQ, part of the EnforceAuth platform, maintains the relationships between the user, the parent agent, the sub-agent, the channel and the alert list. When a sub-agent two hops from the user attempts an action on a resource the user's task never referenced, that distance is a fact the policy can evaluate. ALLOW at hop one does not carry to hop two without re-evaluation.
- Evidence of the drift. Each decision is recorded with the delegation chain, the policy version and the reason, signed, replayable, retained one to four years, exportable to your SIEM. On October 4 the question "why is Security not getting alerts" has an answer that names the sub-agent, its parent, the instruction and the denied or allowed edit.
Zanzibar-style relationship-based access control is on our roadmap; what ships today is the OPA and Rego decision point, Verdict as the authorization MCP server, and EvidenceIQ's context memory, relationships graph and continuous re-evaluation. Our PDP latency figures (p50 under 2ms, p99 under 5ms) are design targets, not measurements.
What this does not do: it does not make the agent interpret "don't send messages" correctly. That is a model problem and it will improve. It makes the misinterpretation unable to silence the alarm.
The agent followed instructions. In 2027 the question is who was allowed to give them, and for what.
EnforceAuth is the runtime authorization layer for AI agents, built on OPA and Rego. See enforceauth.com and read the EvidenceIQ announcement.
About EnforceAuth
EnforceAuth is the OPA/Rego authorization control plane. The customer's OPA decides what people, services, and AI agents may do at action time. Verdict covers high-impact agent actions. EvidenceIQ keeps the record behind every decision.

