We help enterprises say yes to agents.

The future of access control won’t be more static rules, it will be semantic.

Static policy decides what is allowed. a1 notices what is unusual.Illustrative

Problem

Access control has relied on people being lazy.

Most people hold more access than they will ever use. We got away with it, because people do their job and stop.

Agents don’t stop.

Give an agent a task and it will try every path it has until one works. More fine-grained rules won’t change that. They can say an access is permitted. They can’t say whether it makes sense.

How it works

Think of a credit card.

Before you get one, the bank checks who you are and sets a limit. That limit is your blast radius. A few hard rules apply everywhere, like sanctioned countries. Almost nothing you do on an ordinary day touches either one.

What protects you day to day is a third thing. It watches each purchase and asks whether it looks like you.

You used to call your bank before a trip so your card would work abroad. That is more static rules, written by hand. Many banks no longer ask. The third layer got good enough to take the job over.

IllustrativeThe limit caps how bad it can get. It takes a third layer to notice the pattern.

Enterprise access has the first two. a1 is the third.

The same three parts, on a card and in a company
A credit cardEnterprise access
LimitIdentity, credit history, and income decide how much you can spend.IAM roles and entitlements decide how far anyone can reach. This is the blast radius.
Hard rulesSanctioned countries and prohibited merchants, whatever your limit.Explicit denies, and the deterministic policy you write in engines like Cedar and OPA.
Risk layerEach purchase is weighed against how you normally spend.a1. Each access is weighed against the work it belongs to.

A stolen card can usually buy something small. The third layer catches what is unusual. Access should fail the same way.

An unusual pattern is a signal, not proof. Someone legitimate should be able to clear the check and carry on.

The risk layer

A second opinion on every yes.

Static policy answers one question: is this allowed? a1 asks the next one: does it make sense? It looks at who is acting, for whom, on what, and why now.

  • A sentence, not a score.

    Every decision comes with its reasons, in plain words. Your team can read them, question them, and correct them.

  • An answer, not an alert.

    a1 responds while the access is still in flight. It can let it pass, ask for a reason, verify the person, or send it for approval.

  • A question, not a reflex.

    Every prompt takes someone’s time. a1 keeps count of which ones were worth asking, and flags the ones that weren’t.

  • A dial, not a rulebook.

    This is policy too. You just don’t write it rule by rule. Set how cautious to be for each team and resource, start careful, and loosen with evidence.

One access. Four ways to answer.

An agent working for a finance analyst opens vendor bank details it has never touched. Static policy allows it.

Illustrative
Choose how cautious a1 should be

Whatever the setting, static policy holds. A deny stays a deny, and low risk never grants access.

Who gets asked

Approval isn’t one step, or one person. a1 decides who to ask from the risk in front of it.

  1. Automated reviewa1 looks again, with more context. Most questions end here.
  2. A teammateSomeone close to the work.
  3. A managerSomeone accountable for the team.
  4. The resource ownerWhoever answers for the data or system.

Any of them can clear the question or pass it on. None can grant what static policy denies.

After-the-fact review

Seven cases a week.

Everything above happens while an access is in flight. This part happens afterwards. Once a week, a1 looks back over its own decisions and brings people the few it was least sure of, or that carried the most risk. Security and the business owner look at the same case.

Sometimes they disagree. That is the useful part.

a1 puts the two readings side by side, and the case stays open until they have talked. Nothing waits on them. The access was already let through, held, or stopped. What they decide moves the dial for next time. Decision by decision, that becomes a policy of its own: built up over time, not written out in advance.

This week’s seven

Decisions a1 already made this week, brought back for a second look. Static policy allowed every one of them.

Illustrative
01Agentfor a support leadTried eleven routes to the same customer export in four minutes. The ninth was permitted.Held, asked the leadAgreed

Why it was raised

Eight denials and then a success is what searching for a way in looks like. The export itself was inside the agent’s delegated access.

Security’s reading

The ninth route should not exist. Close it.

Owner’s reading Support

The lead did need the export, for a billing dispute.

Outcome

Export released after the lead confirmed it. The ninth route is being removed.

02Employee with agenta recruiterA recruiter asked for one candidate’s interview notes. Their agent also opened pay bands for the whole team.Asked for a reasonDisagree

Why it was raised

The recruiter may read both. Only the notes were part of what they asked for. The agent widened the task on its own.

Security’s reading

Pay bands should need a stated reason every time an agent opens them.

Owner’s reading People

Recruiters check bands before every offer. Asking each time will slow hiring.

Outcome

No agreement yet. They talk on Thursday, and the case stays open until then.

03Agentfor accounts payableRead a vendor’s invoice, then tried to change that vendor’s bank details. Nobody had asked it to.Routed for approvalAgreed

Why it was raised

The instruction came from inside the invoice, not from anyone the agent works for. Static policy lets this agent update vendor records.

Security’s reading

Text an agent reads is not an instruction. This is exactly the change to stop.

Owner’s reading Finance

Declined. Nobody here asked for a bank change.

Outcome

The change did not run. Finance checked the invoice with the vendor through a contact they already had.

04Agentfor the data platform teamA reporting agent that reads one warehouse began reading three, the night after its owner changed teams.Paused two of threeOpen

Why it was raised

The agent’s job did not change. Its owner did, and nobody has taken it over. The two new warehouses hold customer data.

Security’s reading

An agent with no owner should not gain reach. Keep the two paused.

Owner’s reading Data platform

Not yet reviewed.

Outcome

Waiting on a new owner for the agent.

05Employee with agentan engineerAn engineer asked their coding agent to fix a failing test. The agent tried to change a production setting it had only ever read.Routed for approvalAgreed

Why it was raised

The agent holds deploy rights, so static policy allows the change. Nothing in the task called for it.

Security’s reading

Right call. A test fix should never reach production settings.

Owner’s reading Payments

Declined. The setting was correct and the test was wrong.

Outcome

The change did not run. The engineer fixed the test.

06Employeea contractorA contractor’s account read repositories outside the project it was hired for, from a device it had not used before.Asked to verifyAgreed

Why it was raised

A new device, unfamiliar repositories, and a pace faster than a person reads. The verification was never completed.

Security’s reading

Treat the credentials as stolen until the contractor says otherwise.

Owner’s reading Engineering

The contractor is on leave this week. This was not them.

Outcome

Access suspended and credentials reissued. Three repositories were read before the check.

07Employee with agentan on-call engineerAn on-call engineer was asked to confirm their agent’s actions four times during one incident. Every answer was yes.Confirmed four timesAgreed

Why it was raised

Same engineer, same agent, same incident. None of the four checks changed anything, and each one cost time during an outage.

Security’s reading

One confirmation per incident is enough when nothing else changes.

Owner’s reading Infrastructure

Agreed. By the fourth prompt it was being approved without reading.

Outcome

On-call work during a declared incident moves down one setting.

Everything else this week needed no review. The story behind each decision was kept all the same.

Why seven? It is a number people will read every week. You choose yours.

Foundations

Containment first. Then identity. Then judgment.

A risk layer is only as good as what it stands on. First, make sure an agent that goes wrong can’t get far. Then give every actor an identity: the employee, the agent, and the employee an agent is acting for. Then add judgment.

We worked this order out the hard way, across on-prem data centers and several public clouds, thousands of services, and no single control plane. In a place like that the only common ground is the network. So that is where a1 connects: at the proxies and gateways every access already crosses.

Diagram. Three nested layers. Containment is the outer layer. Identity and static policy sit inside it. Judgment, which is where a1 works, sits inside both.
01ContainmentHow far anything can get if it goes wrong.
02Identity and static policyWho is acting, for whom, and what they are allowed to do.
03Judgmenta1Whether this access makes sense right now.

Writing

Notes on securing agents.

  • Oct 2026

    Security has to follow every hop

    Three layers for agent IAM: identity and runtime trust, risk-adaptive authorization, unified enforcement.

  • Oct 2026

    Three questions for agent IAM in production

    Identity and isolation, dynamic authorization, and enforcement across every access path.

  • Aug 2026

    A three-layer model for enterprise agent security

    Harness controls, agentic IAM, and detection and response, and why each needs the others.

  • May 2026

    Agent identity everywhere

    Part one of securing AI agents: identity and provenance across every run surface.

  • May 2026

    The resource side has to protect itself

    Client-side confirmations are a safety layer, not a boundary. The system receiving the request has to know who is acting.

  • Apr 2026

    Gateways are not the security model

    An enforcement point is not the same as deciding what is safe at runtime.

Why the name

In 1775, Lloyd’s Register began classing ships with a letter for the hull and a number for the equipment. The soundest were classed A1.

We write it small, and mean it the same way for security.

Careers

We’re hiring.

a1 is a small team with a hard problem. If it is the one you want to work on, write to us.

  • Founding engineer, infrastructure

    Networking, proxies, gateways, and identity: the layer a1 connects at.

    San Francisco or remoteApply
  • Founding engineer, applied AI

    The models and evaluation behind a1’s reading of access.

    San Francisco or remoteApply
  • Security engineer

    Access control, IAM, and detection, alongside customers’ security teams.

    San Francisco or remoteApply
  • Forward deployed engineer

    Works inside customer environments to integrate a1 and tune it.

    San Francisco or remoteApply

Start by watching.

Book a demo and we’ll walk you through a1 on access like yours. Then run it in shadow mode. It changes nothing, and shows you the cases it would have raised.

Book a demo
  • meng@a1.security
  • conner@a1.security