Agent-Based, Agentless, or Both? What Each One Can Actually See
Agents are not new to vulnerability management. The reason the question keeps resurfacing is that the estate being scanned stopped holding still. A scanner that assumes every host is reachable on a known network at a scheduled time described most organisations reasonably well fifteen years ago. It describes very few of them now, in an estate that includes laptops connecting through a VPN twice a week, cloud instances that live for forty minutes, and OT segments that are deliberately unreachable from anywhere a scanner sits.
So the useful question is not whether agent-based scanning is better than agentless scanning. It is what each architecture can actually see, and which parts of your estate fall outside both.
A note on where we stand before getting to the general case: Greenbone’s scanning-agent capability is currently available as a TechPreview within OPENVAS SCAN (Enterprise only, new license model only), and is being rolled out in a controlled fashion in collaboration with our professional service team as it is validated. The rest of this post is about the architectural question, which is the part worth understanding regardless of whose tooling you run.
What “Agentless” Actually Means
Agentless scanning means the assessment runs somewhere other than the target. A scanner sits at a vantage point on the network and interrogates the host over the host’s own interfaces.
That splits into two quite different things, which are frequently discussed as though they were one:
Unauthenticated scanning looks at the host from outside, with no privileges: open ports, service banners, TLS configuration, exposed applications, behaviour that can be probed. It answers the question an attacker asks.
Credentialed scanning logs in, over SSH, WMI or an API, and reads the system from the inside: installed packages and versions, patch level, configuration.
The distinction matters because it defuses the most common claim in this topic. A credentialed agentless scan and an agent see largely the same host state. Both read local package inventory and configuration. The difference between them is depth of insight (e.g. scanning of every file) and, as the Agent is directly installed on the host, speed.
What an Agent Changes
An agent moves execution onto the host. NIST’s guide to enterprise patch management technologies describes the architecture plainly: “An agent-based patch management technology requires an agent to be running on each host to be patched, with one or more servers that manage the patching process and coordinate with the agents.” The same document is direct about where the approach earns its keep. Agent-based technologies are “strongly preferred for hosts that are not on the local network all the time, such as telecommuter laptops and smartphones.”
Several things follow from that. The host does not need to be reachable from the scanner at the moment the assessment happens; the agent collects locally and reports when it next has a path back. No scanning credentials need to be distributed to, or stored for, every target, which removes a standing privileged-access problem from the network. And reachability stops depending on firewall rules, NAT, or the segmentation between scanner and target.
The practical effect is a shift away from the scan window toward something closer to continuous state.
What Each One Gives Up
This is the half of the topic that vendor material tends to skip, and it is the part a practitioner needs.
Agentless scanning does not assess hosts that are not there. NIST puts the limitation as omitting “hosts not on the local network, such as telecommuter laptops and mobile devices,” and notes that it can be negatively affected by firewalls and network address translation. Short-lived cloud workloads are the modern version of the same problem: a container that exists for the length of a job will never coincide with a nightly scan. Credentialed agentless scanning also needs credentials with real privileges on every target, which is a risk surface of its own and an operational burden that grows with the estate.
Agent-based scanning has a harder structural limit: a substantial part of a typical estate cannot run an agent at all. NIST again notes that hosts “that don’t permit direct administrator access to the operating system, such as many appliances, generally cannot run agents,” and that agents may not be available for every platform. In practice that category covers network and security equipment, storage appliances, printers, most industrial and building-control systems, embedded devices, and anything on the network nobody has administrative control over, which is precisely the group most likely to be neglected already. An agent is also software you have now installed on every host, with its own lifecycle, resource footprint and patching needs.
The deeper limit is perspective rather than reach. An agent looks at the host from inside. It can tell you that a vulnerable package is installed. It is not well placed to tell you that the affected service is reachable from the internet, sitting behind a misconfigured proxy, or listening on an interface nobody intended. That is a question about the network’s view of the host, and it can only be answered from the network.
Why Hybrid Is the Answer That Survives a Real Estate
Laid out this way, hybrid stops looking like a compromise between two options and starts looking like a consequence of the fact that they answer different questions.
Neither architecture’s blind spots are covered by its own strengths, and each covers the other’s. Agent-based scanning reaches exactly the hosts agentless scanning misses, the absent, roaming and ephemeral ones. Agentless scanning reaches exactly what agents cannot, everything that cannot run one, and it supplies the outside-in view that an agent structurally cannot produce. That is an unusual property. Normally one approach dominates and the other becomes legacy.
Standards already assume more than one vantage point. CIS Controls v8 Control 7.5 requires organisations to “conduct both authenticated and unauthenticated scans, using a SCAP-compliant vulnerability scanning tool” against internal enterprise assets. That requirement predates the agent question and is independent of it, but the reasoning is the same: one viewpoint is not enough to describe a host’s exposure.
Hybrid done well does not mean scanning everything twice. It means deciding, per class of asset, which vantage point answers the question you actually have about it.
A Practical Suggestion to Divide It Up
Roaming endpoints and anything that connects intermittently: an agent, where the platform supports one. This is the case for which agentless scanning has no good answer.
Appliances, network and security equipment, OT and embedded devices: agentless, because there is no alternative. These assets also tend to sit in segments a scanner can only reach from a deliberately placed vantage point, which is worth planning rather than discovering.
Servers and long-lived cloud instances: either works, so decide on your credential policy and the rate at which the systems change. If distributing scanning credentials across a large server estate is a problem you would rather not have, agents remove it. If installing software on production hosts is the harder internal conversation, credentialed scanning is mature and well understood.
Anything internet-facing: always also unauthenticated and from outside, whatever else assesses it. This is the only view that reflects what an attacker can reach, and no agent can produce it.
Ephemeral workloads: neither architecture fits comfortably. Assessing the image or template before deployment is usually a better answer than trying to catch the instance while it exists.
None of this makes agentless scanning an incomplete answer to the estate it was built for. A network-resident host assessed with credentials is fully assessed, and the outside-in view of it is still the only one that shows what an attacker can reach. Due to being directly installed on the target host, Agents extend vulnerability management with detailed information on hosts agentless scans can not discover.
Sources
- NIST SP 800-40 Rev. 3, Guide to Enterprise Patch Management Technologies, §4.1.1 and §4.1.2. nvlpubs.nist.gov/nistpubs/specialpublications/nist.sp.800-40r3.pdf
- CIS Controls v8, Control 7.5. cas8.docs.cisecurity.org/en/latest/source/Controls7
