Over 65% of enterprises built custom authorization code. It worked great for human web sessions. But AI agents broke it.
Agents don't click fixed URLs. They:
✔️ Plan dynamic tasks
✔️ Query vector databases
✔️ Invoke MCP tools
✔️ Delegate to sub-agents
When that happens, homegrown role checks lose identity context. You get "Confused Deputy" risks and standing privileges.
Learn how to stop hand-coding authorization into every microservice 👉 hubs.ly/Q04F4hBS0
2
16
A person asks your AI agent one question. But behind that question, dozens of access decisions take place that your login screen never sees.
That's something many organizations are struggling with right now. And that's exactly what we tried to tackle during Agentic IAM Day 2026!
We gathered several security and identity leaders to work through what runtime control actually looks like when the identity performing a task is an AI agent: how to bound what it can reach, govern each action it takes, and prove what happened after the fact.
We also covered:
✔️ Where authorization has to live when one prompt becomes hundreds of downstream calls
✔️ How teams are binding human, non-human, and agent identity under one policy model
✔️ What it takes to move an agent into production with real guardrails, not a six-month review
✔️ Why a fast decision and a wrong decision can be the same decision
Every session is now available on demand, and you can watch them here 👉 hubs.ly/Q04F4kK60
1
2
102
PlainID was mentioned 6 times in Gartner’s October 2026 research 🎉
Every one of those reports makes the same point: the future of agent identity belongs to runtime authorization, under an independent control plane that holds across agents and environments.
For years, identity was checked once, at login. Agents break that model. In seconds, an agent can pull from a vector store, call an MCP server, invoke an API and hand work to a sub-agent, all on a person’s behalf. A role granted at login can’t speak to any of those moments.
Authorization has to happen at runtime, at the moment the agent acts. It needs the full picture: the person it represents, the agent doing the work, the tool it reached for and the data in play.
We’re proud to be part of the conversation shaping how enterprises secure AI agents.
Read the breakdown of each report here: hubs.ly/Q04D-6Rb0
2
43
The PlainID team is heading to the Gartner Identity & Access Management Summit 2026 on December 7-9 in Las Vegas. Come by our booth and let’s talk about the future of authorization for humans, machines, and AI agents.
Book some time to talk with us here: hubs.ly/Q04DRNPC0
#GartnerIAM
20
If you work at a bank and you’re putting AI agents to work, here’s a few things that are worth weighing first:
1. A simple agent that can move money carries far more risk than an advanced model that drafts memos. The OCC’s new Supervisory Framework for Bank Use of Super Intelligence states that control should scale with the consequences of failure and how much you hand over to the agent.
2. Write down each agent’s scope in terms you can test. Who can invoke it, which data and systems it can reach, and which actions it can take. The framework treats activity outside that scope as a control failure, even when the result looks fine.
3. A system prompt shapes behavior, but prompt injection or drift can push an agent past its instructions while its credentials still work. An authorization check that sits outside the model decides what the agent can actually do.
4. An agent can stay online while acting incorrectly. You want to be able to revoke its access across every system it touches from one place, without redeploying anything.
5. When an examiner asks why an agent took an action, you need the recorded inputs, the data it retrieved, the tools it called, and the decision that allowed it.
Michael Anastasio, Regional Sales Manager East at PlainID, each of these to the OCC framework, and to how runtime authorization handles them, in our new blog: hubs.ly/Q04DLdYY0
1
50
Only 15% of organizations are actually governing agent access at runtime.
We ran a live poll during PlainID’s Agentic IAM Day, with security and IAM leaders in the room, and fewer than 1 in 6 can actually govern access at runtime.
Agents are reaching production faster than the controls meant to govern them, which now results in unauthorized access, data the agent should never retrieve, plus activity no one can audit later.
That's why the access decision has to happen at runtime, in context, every single time.
Those agents are in production now. What stays open is whether policy authorizes every action an agent takes, in the moment it happens.
You can watch the full recording of the IAM Agentic Day here: hubs.ly/Q04Dp6Ny0
14
The recording of our virtual summit Agentic IAM Day is now available, and it all comes back to one question: once an AI agent is inside, what is it actually allowed to do?
A perfectly authenticated agent can still take an action the task never called for. Same identity, same permissions, different outcome. That problem ran through every session.
Security and IAM leaders from all around the globe shared their insights with us:
✔️ Where authorization really sits for autonomous agents
✔️ Securing MCP as the new access path
✔️ Why non-human identities need runtime control, not human-speed reviews
✔️ Keeping sensitive data inside the boundary as agents become the ones consuming it
If you missed us last week, now it's your chance to catch up here 👉 hubs.ly/Q04DdKXT0
1
1
48
This month’s Authorized? is out, and it’s all one topic: governing AI agents once they’re actually running.
This edition puts together everything we’ve been working on recently: the runtime side, the build-or-buy question security teams keep hitting, and what the latest breach data says about where this goes wrong.
You can read it here:
hubs.ly/Q04C-bm80
1
29
How do you govern access when the thing asking for it is an autonomous agent, not a person?
We’re bringing a small group of CISOs and identity leaders together in Chicago to talk it through, over dinner, before the holiday rush sets in.
On November 12, PlainID and iC Consult are co-hosting an evening of good food and candid conversation about the access challenges landing on IAM teams right now, how agentic AI is reshaping authorization, and what’s working for the teams already in production.
We skip the panel and the slideware, so you can actually talk. You’ll be in a room of peers comparing notes on the same problems you’re each working through.
Join us here: hubs.ly/Q04CBRnN0
1
3
49
The agents aren’t waiting for your IAM strategy to catch up ⏳ Agentic IAM Day 2026 is tomorrow.
We’re going Beyond Who Gets In, into what your autonomous agents are doing right now and how to govern them with true Zero Standing Privileges.
Join IAM leaders from around the world who already do this in production as they share the playbooks that work and the failures they learned from. Here’s what we’ll cover:
✔️ The new identity crisis: policy sprawl, NHIs, and legacy IAM at its breaking point
✔️ Owning the MCP access path before it owns you
✔️ When agents become the business enabler, and how to scale them safely
✔️ A closing fireside chat on where IAM goes next: Zero Access
There’s one day left to save your spot 👉 hubs.ly/Q04Cgpxb0
21
The assumption we hear most from IAM teams: an AI agent is just another service account.
And it’s this assumption that quietly breaks their whole authorization model.
A service account is one known thing you request, approve, and provision once. An agent collects different credentials for different jobs, spins up for a single task, and disappears. It never asked permission to exist.
So the traditional loop - request it, wait for sign-off, get it provisioned - falls apart. The agents are already running. Most enterprises have several agent registries that don’t even know about each other, scattered across every hyperscaler and framework in the stack.
Which puts one question at the front of any agentic program: who owns this agent? If no one can name an owner, it has no business running inside your enterprise.
Gal Helemski and Andre Priebe, CTO of iC Consult, spent the session discussing exactly that, from discovering the agents you didn’t know you had to holding someone accountable for a fleet of them.
Watch the full recording: hubs.ly/Q04BT7Jy0
2
1
50
AI agents are now the main attacker behind most of 2026’s breaches. Almost none of it needed a new trick: 79% still came down to a stolen key, token or credential.
The finding comes from The State of the NHI & AI Agent Breach Report 2026, released this week by our friends at NHI Mgmt Group. They analysed 200+ real breaches from 2017 to October 2026, across more than 1,000 public sources. A few of the numbers that stood out to us:
✔️ 33 of the 53 breaches they tracked in 2026 involved AI agents, LLM platforms or AI attack routes.
✔️ 44% of all breaches began with a leaked, stolen or weak credential.
✔️ 56% involved more than one kind of non-human identity.
✔️ Supply chain is now the number one known way in, where one stolen token can reach every downstream user.
One leaked key rarely stays one key. A single credential reaches the next secret, then cloud roles, then the data, often without a single alert. And we know these credentials will most likely leak. What decides the damage is what that credential, or a hijacked agent, can reach once it is inside.
NHI Mgmt Group are very clear about the fact that the inventory, scope and short credential lifetimes matter more than any single control.
Controlling the scope and lifetime of access is the work PlainID does. We enforce runtime authorization and Zero Standing Privileges across human, non-human and AI agent identities, so a compromised key reaches only what policy allows at that moment, not everything it inherited.
You can access the full report here 👉 hubs.ly/Q04BT7NK0
42
If you haven’t played Access Creep yet, you should give it a go!
It’s a short game about catching AI agents in disguise before they reach data they shouldn’t have access to. They hide in the flock, carry real credentials, and head for Payments, Customer Data, or HR Records. Your job is to spot them and keep them out, and the rules change on you mid-round.
We call that problem access creep, and it’s what Runtime Authorization handles in production: checking every identity, human or agent, against what it’s actually allowed to reach.
Catch enough of them and you land in the top 10 on the leaderboard, where you can win prizes!
Play here: hubs.ly/Q04B-SnV0
18
Is your IAM strategy built for static human users, or autonomous AI agents?
If you’re still relying on human-speed review cycles to govern machine-speed workflows, you aren’t just behind - you’re vulnerable.
In just 7 days, join us for Agentic IAM Day 2026, a free virtual summit designed to bridge the gap between autonomous innovation and enterprise security.
📅 October 28 | 🌐 Virtual Summit | 🎓 Earn 1 CPE Credit
Save your spot: hubs.ly/Q04BGQ7C0
18
Search for an authorization tool today and the list runs long. So how do you pick the right one?
✔️ One verifies the login.
✔️ One certifies access for the auditor.
✔️ One decides a request inside a single app.
✔️ One enforces what an agent can do while it runs.
They're all good at their job, but they might also solve a problem you did not have, while the gap you were worried about stays open.
So we wrote the guide we wanted to hand people who ask us this. It sorts 16 tools into five clear categories, shows what each one actually covers and where it stops, and gives you a five-question test that narrows a long list fast.
It also covers the tools you already run, because for most teams the real question is not which to replace. It is which layer to add on top.
Read it before your next vendor call 👉 hubs.ly/Q04BHr0p0
1
14
An AI gateway gives you a log of every access request and a way to manage traffic.
But a log doesn’t answer the question that matters in the moment: who is this agent acting for, and is that person allowed to do what the agent is about to do.
The gateway can see the agent, or it can see the human, rarely both on the same request. So it can direct traffic, yet it can’t make an access decision you’d trust.
An agent can pull data or trigger an action in under a second. The window to decide whether to allow it is that short. By the time a quarterly access review runs, the data has already moved.
Authorization has to happen inside that window, on the gateway you already run, tying each agent to the person it’s acting for before the request goes through.
Download our white paper to learn more: hubs.ly/Q04B5-lq0
14
One agent hands a task off to a sub-agent. The sub-agent calls a tool, and the permission from that first handoff no longer matches what the agent is doing.
The old access model decided permissions once, at login, with a coarse yes-or-no scope.
Agents don’t operate that way. They act on their own, across multiple hops, at machine speed, and every hop can carry more privilege than the task in front of it needs.
In our session with @kenhuangus, who co-chairs an AI security working group at the @cloudsa, Kobi Gol from our product team walks through the architecture decision that makes or breaks this at scale, and what he sees enterprise customers doing instead of writing access logic into every agent.
The wider conversation with Ken also gets into:
✔️ Why an agent needs its own identity, separate from the human who triggered it
✔️ How each new sub-agent inherits more trust than it should, and why every hop needs its own check
✔️ How a poisoned MCP tool or agent skill, paired with an over-privileged identity, becomes data exfiltration
✔️ Why the audit trail of who asked, what the system decided, and which policy governed it becomes your accountability record when something goes wrong
Bottom line for the architects and AI platform owners watching: if your access control can’t re-evaluate at each hop, at runtime, with the human identity, the agent identity, the requested action, and the intent in context, it will not hold once agents call other agents.
Watch the full recording here 👉 hubs.ly/Q04B5yC60
1
4
103
We ran a poll during our agentic IAM webinar and asked the participants how many AI agents are currently running inside their organizations. Half of them couldn't answer.
You can’t authorize, govern, or assign accountability to an agent you’ve never discovered. And if nobody can see the agent, nobody owns it, which means nobody is accountable the day it does something it shouldn’t.
Most enterprises already run several disconnected agent registries across hyperscalers, agentic frameworks, and the SaaS tools that now ship their own agents. So before you do anything else, start with visibility: knowing which agents exist, and who owns each one. Policy comes after that.
In the session, PlainID's Gal Helemski and Andre Priebe, CTO of iC Consult, walked through the full agentic IAM architecture across five layers: agent identity, discovery, policy management, enforcement, and audit.
If you’re the IAM or security leader trying to bring agents under control before they multiply, you can watch the recording here: hubs.ly/Q04B5yh30
2
2
49
Login was never the hard part. What an identity does after login is.
In her SuperbCrew interview, our Co-Founder and CPO Gal Helemski explains how PlainID authorizes every human, machine and AI agent in real time, so an agent never gets more access than the person it works for.
Read the full interview here: hubs.ly/Q04B5ykf0
2
31
2 weeks until Agentic IAM Day 2026!
When AI agents outnumber humans in the enterprise, access creep becomes the real risk. It's time to Go Beyond Who Gets In and control what non-human identities can do at runtime.
What to expect on Oct 28:
✔️ Real perspectives from IAM leads at Cisco, Databricks, Thales, Ericsson, Entrust, iC Consult, Humana, Onepoint, NHI Mgmt Group, Radware, EY, and more!
✔️ Practical playbooks for enforcing Zero Standing Privileges
✔️ 1 CPE Credit for live attendees
Don't miss out 👉 hubs.ly/Q04z-FT20
1
1
30











