Skip to main content

Why Cloud Posture Is Not Enough

· 9 min read
Ankush K
Co-founder and CTO, Dwaar
Why Cloud Posture Is Not Enough

Your posture dashboard is clean. Then an exploited CVE lands in the security channel and someone asks the only question that matters that morning: are we running this, and where? The dashboard has nothing to say, because answering that was never its job.

By cloud security posture management, or CSPM, I mean tooling that reads the configuration of your cloud accounts and resources, compares it with a benchmark or policy, and reports what deviates. My position is simple. Keep running it, and stop reading a clean result as a statement about your exposure.

What posture checks earn​

Misconfiguration is a real way to get hurt, and it is not a solved problem. M-Trends 2026 describes attackers exploiting misconfigured Active Directory Certificate Services templates, and deleting backups from cloud storage. The first is a directory setting rather than a cloud one, but the failure has the same shape: a setting permitted something nobody meant to permit. The second raises a configuration question too, because who may delete a backup is a setting, and a settings check is the right tool for reading it.

For third-party cloud accounts, Help Net Security's report on the Verizon 2026 DBIR says weak passwords and permission misconfigurations take close to eight months to fix for half of all findings. That is a third-party figure, and it does not measure posture tooling, but I read it as a sign these settings do not fix themselves, so I would not run a cloud estate without posture checks.

The obvious approach, and where it fails​

The obvious approach is to treat a passing benchmark as cloud security being done. Here is why that breaks, in an illustrative case, not a real incident.

A container service sits behind a load balancer. Its role is least privilege, its storage is encrypted, and the only open port is the one the service exists to serve. The image contains a framework with a remote code execution flaw. A posture check reports clean, and it is correct: nothing about the configuration is wrong.

That is not a contrived flaw. CVE-2025-55182, a remote code execution bug in React Server Components known as React2Shell, went onto the Known Exploited Vulnerabilities catalog on 5 December 2025, added by CISA on evidence of active exploitation. Google's Office of the CISO cites it when saying "Attack speeds can now be measured in days." A flaw like that lives in a package, so it can sit in an account with no bad setting at all.

What the incident data says​

Three incident reports from the past few months rank software exploitation first, though only one of those rankings is specific to cloud:

  • Google's summary of its H1 2026 Threat Horizons report says that in Google Cloud customer environments in the second half of 2025, threat actors exploited software vulnerabilities (44.5%) more often than weak credentials (27.2%).
  • The M-Trends 2026 report says exploits were the most common initial infection vector for the sixth consecutive year, at 32% of the 2025 Mandiant investigations in which an initial infection vector could be identified.
  • Verizon's 2026 DBIR says exploitation of vulnerabilities is now the most common initial access vector for breaches, at 31%, ahead of credential abuse, the previous leader, at 13%. Verizon notes that adding pretexting to its tracked vectors played a role in that fall, and that credential abuse would have been 16% without the change.

These are samples of investigated and reported incidents, not all breaches. They cover different populations, so 44.5%, 32% and 31% do not contradict each other. Only Google's figure above is specific to cloud, and cloud data is not uniform: M-Trends says that in Mandiant's investigations of cloud-related compromises, voice phishing led at 23%, stolen credentials were at 16% and exploits came last of the six vectors it names, at 6%. None of these figures measures posture tools. What I take from them is narrower: a question posture tooling does not set out to answer now sits close to the centre of how many intrusions start, and Mandiant's cloud cases are a reminder that software exploitation is not the only route into a cloud estate.

The questions, one at a time​

Exposure is several questions, and a settings check is built for one of them. The right-hand column is my reasoning from what the category is for, not a measurement of any product.

QuestionEvidence it takesSettings check alone
Is this setting wrong?Configuration and a policyYes, this is the job
What runs there?An inventory of deployed workloads and imagesNo
What is in it?The package list of the artifact actually deployedNo
Is any of it flawed, and is that flaw being exploited?CVE data and CISA's KEV catalogNo
Can an attacker get to the flawed code?Network and identity exposure, joined to the aboveThe exposure half only

The last row is the one people get wrong. A posture check can tell you a port is open or a role is broad. A scanner can tell you a package in an image has a CVE. Neither tells you whether those two facts describe the same running workload. Reachability is a join, and you cannot join against an inventory you do not have, which is why I keep arguing that you cannot secure what you cannot list.

Timing makes the gap worse. M-Trends puts mean time to exploit at "an estimated -7 days" and says "exploitation is routinely occurring before a patch is even released." Help Net Security's Verizon report says only 26% of KEV vulnerabilities were fully remediated, down from 38%, with a median of 43 days against 32 the year before. When the average is negative, patch cadence alone cannot be the whole answer for the flaws exploited first. KEV lags by construction, since a CVE gets listed on evidence that exploitation has already happened. KEV matching is not prevention.

What inventory buys you is a fast, honest answer to "are we running this?" on the day it matters. Posture data then earns its keep a second time: once you know which workload holds the flaw, its exposure settings tell you what to isolate first. KEV does its work as a filter over a long list, which I covered in CVSS Is Not Risk.

Where this position stops applying​

The argument is about a function, not a product label. Tools sold under the posture name may add some vulnerability or reachability context. If yours does, hold it to the table above, and check that it is looking at the artifact you actually deployed.

It weakens as your own code shrinks. If you run a static site on managed services, there is little of your own software to inventory, and configuration and identity are a larger share of what you can get wrong.

It also does not demote identity. Weak credentials are still at 27.2% in Google's data, M-Trends puts stolen credentials at 16% of its cloud-related compromises against 6% for exploits, and the DBIR says that if you count credential abuse at any point in a breach, not only as the first action, it still sits on top at 39%. Posture checks cover some identity settings. A rising share for exploitation does not show posture checks are low value. It shows they answer one part of the problem.

The inventory half of this gap is what Dwaar covers: discovering what you ship from source control, registries and cloud accounts, scanning it, and putting the results in one findings list alongside EPSS and KEV signals.

Try it on your cleanest workload​

Pick the internet-facing workload with the best posture result you have. Without opening the posture tool, write down which build is running right now, which packages are in that build, whether any of their CVEs are in the KEV catalog, and whether the service exposing them can be reached from outside.

Then replay a real date. Had someone asked you on 5 December 2025 whether any running workload included CVE-2025-55182, who could have answered, and from what data? If the honest answer involves asking around in chat, you have found your gap, and no additional posture rule will close it.

References​