PolicyArc AI Agents Get Task-Limited Access

PolicyArc AI agents will no longer need broad, standing access to do useful work. Toronto identity company IDENTOS announced on October 20, 2026, that PolicyArc is now generally available, giving security, platform, and compliance teams a single place to define, enforce, and audit access policy across their applications. PolicyArc AI agents receive a scoped, time-limited version of a person's own access, enough to complete one task and nothing more, with that access automatically revoked when the task or the authorised window ends. The launch addresses one of the fastest-growing problems in enterprise software: agents that pull data, take actions, and make decisions inside company systems were never designed to fit the access controls built for people.

According to IDENTOS, PolicyArc builds on more than 12 years of identity and access work across Canadian government and healthcare systems. The company describes itself as a Canadian identity, privacy, and authorization company, and it says PolicyArc is built on SOC 2 Type II and ISO 27001 certified infrastructure with Canadian data residency. That background matters in regulated sectors, where proving that access was appropriate can matter as much as granting it in the first place.

Why PolicyArc AI agents need a different kind of access

Traditional access control was built around people logging in and staying logged in. A person with a role keeps that role for months, and the system checks credentials at the door rather than at every step. PolicyArc AI agents break that model, because an agent may act dozens or hundreds of times in a single workflow, touching different resources at different stages, and its authority should change as the task changes. According to IDENTOS, most systems only know how to grant broad, standing access, and the company's chief technology officer, Alec Laws, described that approach as reckless when applied to an agent.

The risk is not theoretical. According to IBM and the Ponemon Institute's 2026 Cost of a Data Breach Report, cited by IDENTOS, 92 percent of organizations that experienced an AI-related breach lacked proper AI access controls. An agent with a person's full permissions can read more than the task requires, retain access after the task ends, and act in ways that are difficult to reconstruct afterwards. PolicyArc AI agents are meant to solve that by making delegation narrow, temporary, and recorded from the start.

This is part of a wider shift in how agent infrastructure is being built. Recent launches have focused on giving agents safer execution environments, from hardware-isolated microVMs to lighter sandbox runtimes that confine what an agent can touch. PolicyArc AI agents add the authorization layer to that stack: even a well-sandboxed agent still needs to know which data it may read and which actions it may take, and that decision has to be made in real time, not once at login.

How PolicyArc AI agents get scoped, time-limited access

PolicyArc AI agents work through delegation. A person can grant an agent a scoped, time-limited version of their own access, and that grant is automatically revoked when the task finishes or the authorised period ends. The agent never needs a permanent credential of its own, and it never inherits the full scope of the person's role. If the task changes, the access should change with it, rather than lingering as an open door.

Access decisions in PolicyArc combine standing facts with real-time context, according to IDENTOS. Standing facts include a person's role and whether identity has been verified. Real-time context includes how sensitive the resource is and what stage a workflow has reached. A policy can allow access, deny it, or require extra verification in the moment. That combination is what makes the system a runtime authorization engine rather than a static permission list: the same request can be treated differently depending on the circumstances in which it is made.

Consent is checked on every request, not only at login, according to IDENTOS. Access is checked against what the data owner has actually agreed to, on every access request, rather than being assumed from an earlier sign-in. That distinction matters for agents, which may act long after a person has stepped away from the keyboard. A login that happened hours ago is a poor basis for a sensitive action taken now, and PolicyArc AI agents are designed so that consent does not silently carry forward.

One policy, everywhere, with an audit trail

PolicyArc AI agents also benefit from a simpler management model. Teams can define a resource type once and apply its policy consistently across every connected application, according to IDENTOS. New resource types and rules can be added without rewriting integrations, so access rules can keep pace with the product instead of trailing behind it. For teams managing authorization across a growing product surface, that consistency is a practical advantage: a rule that exists in one application but not another is a gap waiting to be found.

Every access decision is automatically logged and reportable, according to IDENTOS. Audit and compliance evidence exists from day one, instead of being reconstructed after the fact. That matters for compliance and risk leaders who need to prove that access was appropriate, and it matters for incident response, where the first question after a breach is often who, or what, accessed what, and when. PolicyArc AI agents leave that trail by design, because each decision is recorded as it happens.

The intended users are the teams who own access decisions and have to answer for them, according to IDENTOS. That includes platform and security engineering leads managing authorization across multiple products, compliance and risk leaders who need to prove access was appropriate, and teams shipping AI-agent features who need a safe way to delegate access without building that infrastructure themselves. The company says PolicyArc is generally available now, with more detail at its product site.

What PolicyArc AI agents do not solve on their own

Scoped access is a necessary control, but it is not a complete answer to agent safety. PolicyArc AI agents still depend on the quality of the policies written for them. A policy that is too broad, a resource type that is defined too loosely, or a delegation window that is set too long will all weaken the protection, even if the enforcement mechanism works exactly as designed. Authorization is only as good as the rules it enforces, and teams adopting agent features will need to treat policy design as a first-class engineering task, not an afterthought.

There is also the question of how agents behave within the access they are given. A scoped, time-limited grant prevents an agent from reading data outside its task, but it does not guarantee that the agent will use its permitted access wisely. PolicyArc AI agents still need the execution controls, monitoring, and human oversight that other parts of the agent stack provide. The launch is best understood as one layer in a maturing system: sandboxes limit what an agent can touch, offline training limits what a model can reach while it learns, and runtime authorization limits what an agent may do while it works.

Even with those limits, the direction of travel is clear. Enterprises are no longer asking whether agents will act inside their systems, but how to let them act without losing control. PolicyArc AI agents offer a concrete answer to that question: give the agent exactly the permission it needs for one task, nothing more, and make the access disappear when the task ends, with every decision on record. If that model becomes standard, the era of agents running on a person's permanent credentials may come to look like a brief and risky experiment.

Sources: Business Wire, IDENTOS Launches PolicyArc; ID Tech Wire, IDENTOS Launches PolicyArc for Runtime Access Control and AI Agents; IDENTOS, IDENTOS launches PolicyArc. Related coverage on this site: Google Agent Compute Engine Gives Each AI Agent Its Own MicroVM; Anthropic Sandbox Runtime Gives AI Agents a Lighter Security Cage.