packets.buildIn some company of sixty people, today, a CTO is answering a customer’s security questionnaire. Do you encrypt backups? Do you review who has access every quarter? The CTO answers from the policy, because the policy is the only place the answer is written down.
The CTO is not lying. The policy was true the week it was written. Nobody has checked since.
In Zero to One, Peter Thiel asks a contrarian question: what important truth do very few people agree with you on? Ours is this: a company is as secure as what is true about it, not as what is written about it. Almost nobody checks the difference.
The industry asks other questions. Is the policy written? Is it approved? Is the file uploaded?
We ask rather a simple one: is it true?
01 · The mistake
Three jobs, three silos
In a growing company, one person writes the policy. Often it is the CTO, often from a template. The engineers do the work in the way that works that week. Someone else, a GRC hire, a consultant or a tool, collects what the company hands over and files it as proof. These are three jobs, done in three places, at different times.
Every generation of tools made one of these silos better and left the others alone.
- Spreadsheetsimproved the paperwork.
They gave the policy one place to live.
- GRC platformsimproved the tracking.
They gave tasks owners and gave files a place to be collected.
- AIimproved the writing.
It drafts the policy, the risk register and the questionnaire answers.
Each silo got better at its own job. Each tool is judged only on that job: was the policy written, was the task closed, was the file uploaded.
- Was the policy written?
- Was the task closed?
- Was the file uploaded?
- Do the policy, the work and the proof describe the same company?Nobody’s job
So the gap between them stayed where it was.
Size does not fix it
Even large enterprises have not solved this. An enterprise has a team for each job.
- One group
Writes the policy and describes how things should be.
- One group
Builds and runs the systems, and chooses what works in its own stack.
- One group
Does the compliance work, and judges what it can see, yes or no, from the evidence it is given.
They use different words, they are measured on different things, and they rarely sit in the same room. Many do not know the language the others speak. More people did not close the gap. They gave each job its own team, and more distance between them.
02 · The trap
Make-believe is easy
Companies fall into a trap. They come to believe their security is better than it is. They do it in two ways. They write policies that describe the company they want to be. Or they collect evidence that is only partly correct.
Why do capable teams do this? Usually because of time. A security questionnaire arrives, and an enterprise deal depends on the answers by Friday. A customer asks for a certification the company does not have yet. There is no compliance team, and there is one week. The fastest way through is to write the policy that matches the answers, and to collect whatever evidence already exists. This is not dishonesty. It is what the deadline asks for. The policy and the evidence get made for the deadline, not for the company.
In security, the devil is in the details. A screenshot shows that backups are switched on. Nothing shows that anyone ever restored one. A policy forbids credentials in code, and one configuration file still holds one. The paper says one thing, and a small detail says another.
The paper saysA screenshot shows backups are switched on.
A small detail saysNothing shows that anyone ever restored one.
The paper saysThe policy forbids credentials in code.
A small detail saysOne configuration file still holds one.
03 · The certificate
A certificate does not settle it
A certificate says someone checked a sample of a past period. That is useful, but it cannot see every detail.
04 · Why now
AI changed three things at once
1
Attackers use it, and they are faster.
CrowdStrike’s 2026 threat report says attacks by AI-enabled adversaries rose 89%. The average time from first access to spreading inside a network was 29 minutes. The fastest was 27 seconds.
29 minaverage, first access to spreading
27 secfastest
CERT-In reports AI agents getting into Indian organisations using credentials left in configuration files, public code repositories and cloud roles with too many permissions. None of the organisations noticed. In October, Binance lost about $6.3 billion. Its first finding was a compromised engineer’s laptop. It later reported more points of breach and has not published a full account. We think attacking costs almost nothing now, and every company has weak points.
2
Text became free.
A model can write a policy, a risk register and a questionnaire answer in seconds. Make-believe has never been cheaper to produce, so paper no longer proves anything.
3
The checking became possible.
Comparing a policy with real systems, line by line, every week, was too much work for a growing company to do by hand, so nobody did it. Autonomous agents can now read the configuration, ask the owner for the file, and compare. Checking that was too costly to do often can now be done all the time. The same advance that lets an attacker try every door lets a defender check every door.
That is why this can be built now, and why it should be.
05 · What we believe
How we close the gap
The policy, the work and the proof usually live in three places, with three owners, written at three different times. We keep them together on one control, with one person accountable. When one of them changes, we see it. That is how the gap gets closed, and how we find out when it opens again.
There are five steps.
Say what good looks like for this company.
We start from your ideal, not a template: what your company should do, given what you build and what you hold.
Measure the gap.
We compare the ideal with what your systems do today, in the details. The answer is not pass or fail. It is partly there, there but untested, or there and working. For each gap, we say what could go wrong because of it.
Close it, in order.
The most important controls come first, and they get tighter as the business grows. A named person is accountable. Autonomous agents do most of the work, and a check in the system confirms the work is really done.
People decide what matters: whether a control applies, whether a policy is approved, whether a risk is accepted, whether the plan goes ahead, and whether a stage is attested. We tell you how long each step takes, including when a date cannot be met.
Keep it closed.
A control is done when it works, not when the task is closed. We count it as complete when evidence arrives on schedule, because the control is running. When the evidence stops, the gap has opened again, and we reopen the work.
Have someone else check.
When we attest that a company has reached a stage, an independent assessor does the review. We engage that assessor, and we say so wherever the attestation appears. We say “attested”, never “certified”.
An attestation is not a certificate. A certificate is issued once and describes one moment. Our attestations are continuous: an attestation is a live representation of where the company stands now, and it changes when the controls or the evidence change.
- Partly there
- There but untested
- There and working
How the autonomous agents work
Autonomous agents are part of how packets works, not an extra. This is what we ask of them.
- Whether a control applies
- Whether a policy is approved
- Whether a risk is accepted
- Whether the plan goes ahead
- Whether a stage is attested
- 01
Autonomous agents do the work, not just track it.
A tracker tells you what is missing and leaves the work to you. Our autonomous agents do the work. They read your systems, website and documents to pre-fill the first answers. They ask owners for files, remind them, and say so when someone goes quiet. They open tasks, and they draft questionnaire answers from your evidence. They check that the evidence matches the policy, not only that a file arrived.
- 02
Every claim has a source.
Nothing in packets appears without where it came from: a setting, a file, or the person who said it. Everything an agent finds is marked “not confirmed” until a person confirms it. If we find nothing, we show “no signal found”. We never fill the blank, and silence never counts as a pass.
- 03
Autonomous agents write. Checks and people decide.
An agent can draft a policy, a risk, an answer or a plan. It cannot mark its own work finished or approve anything. A check in the system confirms the work against the facts, and people decide the five things that matter. A person reviews every questionnaire answer before it goes out, and approves the plan, including any date that cannot be met. If an agent cannot do a step, a named person does it.
We do the implementation. The audit opinion belongs to the auditor.
A week with packets
Here is how we have designed a week to run.
MonRead
Autonomous agents read your cloud setup, your code repositories and your identity settings, and compare them with the policy. They list where the two differ.
TueWalk through
Your named person walks your CTO through each difference and what could go wrong because of it.
WedDecide
The CTO decides what to fix first. Autonomous agents open the tasks, with owners and dates.
ThuCollect
For each control that should be working, autonomous agents ask the owner for the evidence, remind them, and escalate if they go quiet. They check that the evidence matches the policy, not only that a file arrived.
FriReport
One short report: what works, what has no evidence this week, and what changed.