What it covers
The host agent watches a set of directories you configure (by default your documents, desktop, and pictures folders) for file changes, and reasons about what it sees rather than matching a known ransomware strain by name.
| Capability | What it does |
|---|---|
| File-integrity monitoring | Fingerprints the watched directories into a baseline, then watches for changes: file content, mass renaming, and deletion, reported against configurable thresholds for how much activity in how short a window is worth an alert. |
| Ransomware detection | No strain has to be known in advance. Instead, several signals that each have an innocent explanation on their own (content becoming high-entropy, names changing in bulk, backup files disappearing, a note-like filename appearing, honeypot files being touched) are combined into named, calibrated conjunctions. An archive tool compressing a folder also produces a sharp entropy jump, for instance, but it does not also rename the files it touches; that distinction, and others like it, is what keeps the detector from crying wolf on ordinary tools. |
| Honeypot files | Decoy files placed in watched directories that no legitimate process should ever touch. Any modification to one is treated as a strong signal on its own. |
| Process attribution | When an alert fires, the agent tries to identify which process is responsible by scanning for whoever currently holds files open in the affected directory, attributed to the burst of activity as a whole, since the specific file that triggered the alert is very often already closed by the time the alert is evaluated. |
| Containment | Killing the responsible process and isolating the host's network traffic. Both are refused unless containment has been explicitly armed on that host, and network isolation is reversible: it carries an undo record and, by default, a time limit, so a false positive costs minutes of connectivity rather than a site visit. Sold and consented as a separate add-on from this plan — see Host containment for what it does and the two conditions required before it can act. |
| Host log ingestion | journald and syslog records on Linux hosts, filtered by the shape of the record rather than by raw volume: a curated set of identifiers and message patterns, because an unfiltered host log is overwhelmingly noise from services that have nothing to do with security. Windows event logs are not ingested. |
| Malware analysis | Signature-based scanning is only as good as its signature database's age, and a scan run against a stale database is reported as unknown rather than clean. A separate, declared subset of YARA rule syntax is supported; anything outside that subset is refused by name at load time rather than silently dropped, and its output is never presented as full YARA. |
| Application control (observed only) | Reads what a host's own allow-listing (where present) has already decided and where it would have to hold and does not. This never blocks anything itself. A result may say a host is unconfined, and must never say something was blocked, because nothing here can block it. |
| Backup integrity | A five-rung ladder, from a backup simply existing up to its format's own integrity check passing across the whole archive, because a backup nobody has actually tried to restore is indistinguishable from no backup at all. A format that cannot be checked (an uncompressed archive with no built-in checksum, for instance) is reported as unable to reach the top rung, rather than credited with it. |
Containment requires its own consent
Enrolling a host for detection does not by itself enable containment. Arming it — kill, isolate, or both — requires a separate subscription, a separate typed consent, and the host operator's own literal set in the agent's environment. See Host containment.
What it stops short of
- Containment is refused by default on every host. Killing a process or isolating network traffic only happens once an operator has explicitly armed that capability for that host; the agent will alert without it, but it will not act.
- Process attribution works by finding whoever holds a file open in the affected directory at the moment the alert fires; a process that already opened, wrote, and closed its files is no longer visible by the time detection catches up, so attribution can name the wrong process or none at all during a fast-moving burst.
- Windows event logs are not ingested by the host log pipeline: only journald and syslog on Linux hosts.
- Malware scanning is only as current as its signature database; a scan against a stale database is reported as unknown, never as a clean result.
- The declared YARA-subset engine is not YARA. A rule using syntax outside the declared subset is refused at load time rather than partially honored.
- Application-control checks observe an existing allow-list; they never enforce one, and a host with no allow-list configured is reported as unconfined rather than protected.
- An archive format with no built-in integrity check (an uncompressed tar, for instance) can never be verified past the rung where its contents were merely extracted; the report says so rather than rounding up.
How to run it
- 1
Configure the agent
Set the directories to watch, and, if you want containment available at all, arm it explicitly for that host. - 2
Build a baseline
The agent fingerprints the watched directories once before it starts watching for changes. - 3
Enrol and start watching
The agent reports its status and events back to your account, visible from the host agent page. - 4
Review events as they arrive
A CRITICAL alert captures available forensic evidence before any containment action is even considered, so evidence is preserved whether or not containment is armed on that host.
Read before relying on any result
Sentinelleai is an automated security monitoring platform. Nothing it produces is an audit, a penetration test, a compliance opinion, or a guarantee that a system is secure. "No findings" means the tools that ran found nothing, never that nothing is there; anything a tool could not check is reported as unknown, not as clean. For decisions that carry real money or real liability, engage qualified professionals and treat Sentinelleai's output as their starting material.