Skip to main content

You Cannot Secure What You Cannot List

· 9 min read
Ankush K
Co-founder and CTO, Dwaar
You Cannot Secure What You Cannot List

A new CVE lands and someone asks the question that decides your next few hours: do we run this, and where? If the answer comes from a spreadsheet, you are not answering that question. You are reporting what somebody wrote down the last time they opened the spreadsheet.

The list everyone starts with​

The obvious approach is to write it down: a spreadsheet, a CMDB, a periodic request asking owners to confirm their rows. It feels responsible, and it gives an auditor something to read.

My position is that a list kept by hand is always wrong. I have no study measuring how wrong, so this is reasoning and not a statistic. Systems change when somebody pushes code, publishes an image or creates an account. The list changes when somebody remembers. Nothing connects those two events, and the drift runs in the worst direction: a wrong row gets noticed because someone trips over it, while a missing row looks exactly like nothing.

That is why "we scan everything" is a dangerous sentence. A scan pointed at the list reports on the list. It comes back clean on what it was given and silent about the rest.

Equifax is worth keeping in mind, though it is not a clean inventory story. In 2017 an Apache Struts flaw on a consumer dispute portal was not found during company-wide patching, and attackers accessed personal information of at least 145.5 million people. GAO reports that Equifax's own investigation named four major factors: identification, detection, segmenting of access to databases, and data governance. On identification, the full report gives the account of Equifax officials: the US-CERT notice went to a recipient list that was out of date, so the people responsible for installing the patch did not receive it, and a network scan a week after the flaw was identified did not detect it on the portal. GAO does not say the portal was missing from an inventory. My reading is narrower: an out-of-date list decided who heard about the fix, and a scan that did not report the flaw was no proof it was absent.

What the standards actually ask for​

CISA's BOD 23-01 calls asset visibility "a basic pre-condition for any organization to effectively manage cybersecurity risk". CIS Control 1 wants organizations "to accurately know the totality of assets". NIST CM-8 names the failure: "lack of accountability that occurs when component ownership and system association is not known".

Everyone agrees with those sentences, so they change little. The useful part of BOD 23-01 is the mechanism. It requires federal agencies to run automated asset discovery every 7 days, covering at minimum the whole agency IPv4 space, and vulnerability enumeration "across all discovered assets" every 14 days.

Read that order again. Enumeration runs across what discovery found, and the up-to-date inventory appears among the outcomes the directive wants, "without prescribing how to do so". The idea I take from it: discovery should decide what the inventory says, not the other way round. That is my reading and not CISA's rule. CISA's implementation guidance says inventories "from asset management solutions" may feed into the results of a discovery effort, and it uses authority-to-operate inventories to illustrate which cloud assets agencies must discover. So a list can feed discovery. What I object to is a list standing in for it.

Why a network sweep is not enough​

You cannot copy the mechanism wholesale. The directive covers IP-addressable assets and says plainly: "The scope excludes ephemeral assets, such as containers and third-party-managed software as a service (SaaS) solutions." That is a fair scope for federal networks. It also leaves out much of what an engineering team ships.

So keep the principle and change where you ask. None of the standards above tell you to discover from source control, registries and cloud accounts. This is my extension of CISA's logic: ask the system that holds the thing, because it can answer from its own records without anyone remembering anything.

Ask this systemWhat it knowsWhat a hand-kept list loses
Source controlEvery repository in the organization, touched recently or notForgotten repos that someone may still build from
RegistriesWhat was actually published, tag by tagShadow images, pushed by a job or a person and never written down
Cloud accountsWhich accounts exist and what runs in themUnowned accounts, such as the test account whose owner moved on

The payoff is in comparing them. Source control says what you could build, registries say what you packaged, and cloud accounts say what you run. Their disagreements are the most useful thing you will get. An image with no repository behind it is one you cannot explain by reading code. A repository nothing has been built from is either dead, so archive it, or shipping through a path you do not know about. A hand-kept list holds a single view, so it cannot disagree with itself.

What falls through​

Two well documented intrusions show the shape, with a caveat. Microsoft disclosed that Midnight Blizzard used password spraying to compromise "a legacy, non-production test tenant account" with no MFA. Mandiant reported that in the 2024 intrusions into Snowflake customer instances, credentials from infostealer output "were still valid, in some cases years after they were stolen, and had not been rotated or updated".

Those are credential and MFA failures first. The link to inventory is my reading: a control review only reaches accounts somebody knows to review.

The exposure side is better measured. M-Trends 2026 has exploits as the top initial infection vector for the sixth consecutive year, at 32% of intrusions. Verizon's 2025 DBIR has exploitation at 20% of breaches, up 34%. The figures differ because they count different things, so do not average them. Neither measures inventory failure. They show that exploitation is a leading way in. The step to inventory is my reasoning: a system nobody listed is easy to skip when the patches go out.

Where this stops applying​

Discovery gives you a list, not a decision. I have written about what has to sit between a scan and a decision and how to rank what comes out. CM-8 lists software owners among the data an inventory entry needs, and a discovered asset does not always tell you who answers for it. That part stays human work.

Discovery also sees only what it is connected to. A cloud account nobody connected stays invisible, so one short hand-kept list survives: the roots you ask. None of the three systems will tell you which SaaS your teams signed up for, either.

Completeness is not on offer. BOD 23-01 itself concedes that "achieving full vulnerability discovery on the entire enterprise may not complete in 14 days", and it has agencies track coverage. Do the same: report coverage as a number with a date.

And if you have one repository, one registry and one cloud account, a single page will do, right up to the second account.

Keeping an inventory in step with what you actually ship is the problem Dwaar works on: discovering what you ship from source control, registries and cloud accounts, scanning it, and putting the results in one findings list with EPSS and KEV signals alongside.

Run the diff on one registry​

Pick one registry your team pushes to. Get its full list of repositories and tags from the registry itself, through its own API or console. Put that next to your current inventory and count three things: what the registry holds that your inventory does not, what your inventory claims that the registry does not hold, and how many entries have a named owner. Write those numbers down with the date. Run it again in 7 days, borrowing CISA's cadence, and see whether the first number moved without anyone telling you.

References​