Security Engineering & Tooling · Chapter 3 of 8
Cloud security — CNAPP, agents, and the alert-volume problem
What actually merged and what was bolted on; the real agentless-versus-agent trade-off; and which three things genuinely reduce ten thousand findings.
CNAPP claims the union of cloud posture, workload protection, entitlements, application posture and increasingly data, Kubernetes and detection. Some of that merger is real engineering and some of it is a slide. Knowing which half is which is the whole answer.
The structural critique is fair and widely made: most of these platforms did not start unified. They are the result of multiple acquisitions, each originally designed for a separate function — the canonical example was assembled from five companies, and its vendor then spent over a year re-platforming it precisely because of “fragmented user experiences and operational inefficiencies due to multiple acquired technologies.” That re-platforming is real, expensive engineering; integration that is only co-selling is not.
A currency check: Prisma Cloud is being retired as a brand. In February 2025 Palo Alto announced Cortex Cloud as its next version, merging CNAPP with cloud detection and response on the Cortex platform. Saying “Prisma Cloud” is not wrong; knowing it is now Cortex Cloud signals you are current.
Agentless versus agent-based — the real trade-off
The marketing framing — agentless means no friction, agents are legacy — is wrong. The real trade-off is temporal, not operational.
| Agentless (snapshot / side-scanning) | Agent-based (eBPF or kernel module) | |
|---|---|---|
| How it works | Takes a cloud-provider snapshot of a disk volume and analyses it out of band via API | Runs inside the workload, observing execution |
| Genuinely wins | Total coverage in days not quarters. No in-guest CPU cost. No kernel module to crash your fleet. Reaches machines nobody will let you install software on. Discovers the workloads you did not know existed — which is the actual first problem | Runtime execution visibility. Detection during an attack. The ability to block. The runtime data that makes reachability real |
| Structurally cannot | See anything that exists only in memory. Its view is a point-in-time photograph of a disk — so it cannot detect a fileless attack, a reverse shell, or in-memory credential theft, because none of those touch the volume it snapshotted | Achieve complete deployment coverage, ever. And an agent with kernel privilege is a fleet-wide availability risk — your security tool becomes a single point of failure for every machine it protects |
| Under-discussed cost | Snapshot storage, cross-AZ transfer and scanner compute are a real bill. It is not free, it is billed to the cloud account instead of the workload | Resource overhead in the guest, and an upgrade treadmill on every host |
The correct answer: they solve different problems and the industry converged on both. Agentless for coverage and posture — find everything, know what is misconfigured. Agent or eBPF for runtime detection and response on the subset that matters. The consensus reference architecture is now explicitly hybrid. Anyone selling you one as a replacement for the other is selling.
Ten thousand findings — what actually reduces them
Genuinely works. 1 — Attack-path and graph context, collapsing ten thousand findings into perhaps fifty exploitable chains, so the unit of work becomes the path rather than the finding. This is the highest-leverage reduction available. 2 — Reachability and runtime filtering, which removes most container CVE noise because base images ship enormous amounts of code that never executes. 3 — Ownership routing. Unattributed findings are never fixed; mapping a finding to the team that owns the resource, via tags, IaC provenance or code owners, converts a security backlog into engineering tickets.
And two things that only appear to work.
Severity filtering alone. “Only show criticals” discards exploitable mediums and keeps unexploitable criticals. CVSS was never a risk score.
Auto-suppression rules. They reduce the count without reducing the risk, and they quietly accumulate permanent blind spots that nobody revisits.
If a programme's finding count halved, the first question is which of these five happened.
Check yourself
What genuinely merged in CNAPP?
CSPM, CWPP and CIEM — they share one substrate: cloud API access plus a disk snapshot. The real product is the graph that chains exposure to workload to identity to data.What is thin in CNAPP, and why?
ASPM and code scanning — roughly 12–18 months behind standalone vendors. Structurally: the buyer is the cloud team, the user would be a developer, and the console is not in a developer's workflow.What can agentless scanning never see?
Anything that exists only in memory. It is a point-in-time photograph of a disk — so fileless attacks, reverse shells and in-memory credential theft are invisible to it.What is the under-priced risk of an agent?
Kernel privilege makes your security tool a fleet-wide availability risk. July 2024 was the demonstration.Name three things that really cut finding volume.
Attack-path context, reachability/runtime filtering, and ownership routing. Severity filtering and auto-suppression only look like they work.Is there a Gartner MQ for CNAPP?
No — a Market Guide, most recently August 2025. Anyone citing an MQ for CNAPP is citing a different market's document.Companion chapter: this one is about the tooling market. For the failure modes themselves — why cloud breaches are overwhelmingly customer misconfiguration rather than provider compromise, and the small set that recurs — read Cloud security.