DPLens runs on the machines that matter most, with access to their logs. This page sets out what it does, what it deliberately does not do, and how to tell us if we have got something wrong.
It does not phone home
- No telemetry. DPLens sends nothing about you, your machines or your events to us. There is no usage reporting and nothing to opt out of.
- No licensing service. Licence keys are verified offline against a key built into the program. An air-gapped machine is a normal deployment.
- No auto-update from us. DPLens never downloads anything from DPLens. You decide when to upgrade — by running a package you fetched and verified, or, if you run your own DPLens Manager, by asking it to upgrade the agents enrolled with it, which then verify the package themselves (see Automatic upgrades from the Manager).
- No cloud service and no control plane of ours. The agent is standalone. The optional DPLens Manager is a server you run yourself, and an agent connects to it only after you enrol that agent.
The only outbound connections DPLens makes are to the destinations you configure; to your own DPLens Manager, if you have enrolled the agent with one; and, if you switch it on, the public-IP enrichment lookup. That last one contacts a third-party address-reflection service, is off by default, and lets you point it at any endpoint you prefer.
The shape of the process
DPLens runs as two processes, not one — three on an agent enrolled with DPLens Manager.
The collection service does the work, as a virtual service account with no password and only the privileges it needs.
The console host is separate, with fewer privileges. It cannot read your event logs, reach your destinations, or read the secret store; it asks the service and renders the answers. A fault in the web interface cannot become a fault in collection.
The Manager connection — only on an agent you have enrolled with your own DPLens Manager — is a third process, under its own local account, confined so that it cannot start other programs and capped in memory. It makes the outbound connection and nothing else: it holds no key, cannot read the configuration or the secret store, sees the enrolment token only while it relays the enrolment request, and asks the service — at a limited rate — to sign the proofs of identity a connection needs. An agent that has never been enrolled has no such process and no such account.
See The service and its permissions.
Data at rest
| What | Where | Protection |
|---|---|---|
| Configuration | config\agent.yaml | Administrators and the service only |
| Licence | config\licence.key | As above. Signed, not secret. |
| Secrets | state\secrets\ | Sealed to this machine. Never displayed or written out in clear. The one way out is the deployment capture, which the agent performs itself and only re-encrypted under a deployment key that is shown once and never stored — see The deployment MSI. |
| Undelivered events | state\cache\<destination>\ | As above — this is your log data, so the folder permissions matter |
| Audit log | state\audit\audit.jsonl | Append-only, hash-chained |
| Console password | state\secrets\ | Stored as a hash, sealed like any other secret |
| Sign-in lockout state | state\auth\ | As above |
| Manager enrolment (enrolled agents only) | inside the state folder | As above. Contains no secret: Manager-delivered secrets are in the secret store, and the agent's identity keys are held by Windows. |
| Agent identity keys (enrolled agents only) | Windows key storage — in the machine's security hardware where it has it | Cannot be exported from the machine, by the agent or anyone else |
Permissions are applied at installation and re-applied and verified at every service start, so a change made by something else is corrected rather than inherited.
Secrets are write-only
You can set a secret and see that a handle is in use. Nothing shows you what is behind it — not the console, not a log, not the configuration file, not an export.
Sealing is to the machine: copying the state folder onto a different, separately-installed machine does not carry the values with it. That is what protects a stolen or mislaid copy of the folder.
The same holds for every credential a destination uses — a Splunk token, a ClickStack ingestion key, a client certificate's key. The configuration carries only the handle it is sealed under. A ClickStack destination's auth_header setting is the name of the header that carries the key, never the key itself; the key is not shown in the live stream, and if a collector echoes it back in an error message it is removed before that message is displayed or logged.
Sealing is not a boundary between accounts on the same machine, and it is not a substitute for keeping secrets out of an image. The folder permissions are what keep other local users out, and a machine cloned from an image can carry usable material with it — so do not bake secrets into a golden image. See Golden images and VDI clones.
Data in transit
- To your SIEM: TLS with certificate validation on by default, and mutual TLS where your receiver requires it. Validation can be turned off for testing and says plainly what that costs.
- To ClickStack and other OpenTelemetry collectors: HTTPS is supported, with the same certificate validation and mutual TLS. It is off by default, because a self-hosted ClickStack collector listens on plain HTTP — so until you turn it on, your events and the ingestion key cross the network unencrypted. The console marks such a destination key sent unencrypted. See Sending to ClickStack.
- To the console: HTTPS. Loopback only unless you turn on remote access, which requires an explicit list of permitted networks. That list is checked before the TLS handshake.
- Masking runs before anything leaves. A value a masking rule rewrites is not written to the disk cache and not sent, and the console's own live stream shows the rewritten form, because that is what leaves.
- To your DPLens Manager (enrolled agents only): outbound only, TLS 1.3 only, to the address you installed, and only after the server proves its certificate comes from the authority you pinned. The agent proves its own identity with a key that cannot leave the machine. Inspecting proxies are refused rather than trusted.
What a managed agent sends
An agent enrolled with your own DPLens Manager sends it the following — to a server you run, not to us. The message format has no field that can carry free text: every value is a number, a hash, a fixed code or a strictly formatted identifier, and the agent's build refuses to compile a field that could carry anything else.
When it enrols, once:
- the enrolment token, and the public halves of its two identity keys, with where they are held (security hardware or Windows software key storage);
- the DPLens version and the configuration format version it understands;
- one-way digests of the multiple unique hardware and software identifiers its licence is locked to — never the identifiers themselves;
- the host facts listed below, and a fingerprint of its running configuration;
- a signature proving the keys belong to this connection.
At every check-in (every one to five minutes):
- its version and build, a random identifier for this run of the service, a sequence number, its clock reading and how long it has been running;
- its position in the Manager's sequence of commands;
- fingerprints of the configuration it is running and the one it was assigned, the Manager's version number for it, and whether it has drifted — and if so whether by a file edit, a local change or a start-up refusal;
- its licence state: which licence decides, the expiry stage, any hold and its reason, the lease identifier and expiry, and whether a local key is present;
- whether it has asked to leave, and whether anything below was cut short for space.
Every five minutes, as a summary of the window: counts of events collected, delivered, deliberately filtered, lost, aggregated and masked; volumes in and out, and an estimate of what field pruning saved; disk cache depth; peak event rate; the agent's own memory and processor use; failovers and restarts; the fullest cache; how many destinations are not delivering and how many components are unhealthy; and whether collection was running or held, spilling to disk, upgrading or managed in that window.
As counter changes: the same numeric counters the console shows — per processing stage, for the sources, per destination, for the caches, failover, the licence, component restarts and the Manager connection itself — each keyed by a fixed number, never a name.
Only when something changes:
- host facts: the Windows version and edition type (server, client or core), the computer name, the Active Directory domain, up to eight IP addresses, the identifier digests above, where the keys are held, which kind of account the service runs as, whether the state folder's permissions are intact, whether the program's own signature verified, which kinds of source are in use, which recommendation rules for installed services apply, and whether the Manager may configure receivers;
- component health: each component's kind and position in the configuration, its state, how it last stopped, a fixed error code and its restart counts;
- destination health: each destination's position, delivery state, fault class, cache fill and depth, which of its addresses is active, how many addresses it has, and its failovers;
- recommendations: the identifiers of the rules that apply, and how many findings each has;
- the audit log: whether its hash chain verifies, and how many records of each class — configuration, sign-in, secret write, licence and so on — were written since the last report;
- the names of the Manager-delivered secret handles the agent holds;
- upgrade progress.
When it finishes a command from the Manager: the result code; for a configuration, fingerprints, whether collection is running or held, the names of any Manager secrets it was missing and — if it refused the configuration — the position of each problem as numbers (which section, which entry, which setting), never a message; for an upgrade, the versions involved and the reason code; for a licence, the lease identifier; for a detach, the names of the Manager-delivered secrets it turned into local ones; for a renewed certificate, the fingerprint of its unchanged public key.
Only if the Manager asks after restoring from a backup: a copy of the last signed command and licence lease it holds, so the Manager can prove where it left off.
What a managed agent never sends
- Event content — no log line, message, field value or sample of what it collects.
- Configuration text — only fingerprints. The Manager already holds what it rendered; it never sees what you set locally.
- Free text of any kind — no error message, component name, file path, destination address, destination host name, user name or command line.
- Secret values — not the credentials the Manager sent, not your own, not the console password.
- The hardware and software identifiers themselves — only digests.
- Audit record content — only counts per class and whether the chain verifies.
- Anything to DPLens. All of the above goes only to the Manager you enrolled the agent with.
How it is built
- A memory-safe language, with no unsafe code outside the small, reviewed layer that talks to Windows.
- Every parser, decoder and encoder is fuzz-tested. Anything that reads untrusted input — a syslog message, a NetFlow packet, a log line, a configuration file — gets a fuzzing target in the same change that adds it.
- Patterns you write run on an engine that does not backtrack, so a pattern cannot be written, by accident or otherwise, that stalls the pipeline.
- Bounded by construction. No unbounded queue exists anywhere. Sources are pull-based and pausable, so pressure travels back to the source rather than accumulating in memory.
- No dynamic code loading of any kind. Nothing is downloaded, plugged in or evaluated at runtime.
- A deliberately small dependency list. Every third-party crate has to justify itself against writing the code ourselves, and its licence is checked automatically.
Signing and the bill of materials
Released artifacts are Authenticode-signed and published with SHA-256 checksums and a software bill of materials in both CycloneDX and SPDX formats.
Verify what you downloaded before you run it — see Verifying a download, which also explains what an unsigned pre-release means and how to tell.
The agent checks its own signature at start-up and records what it found in the audit log.
The audit log
Every change is recorded: sign-ins and failures, configuration applies and rollbacks, password changes, certificate installation, licence changes, masking policy changes, and the agent's own start-up checks.
Each record is chained to the one before it with a cryptographic hash, so a record cannot be altered or removed without breaking the chain, and the console shows whether the chain verifies.
It is append-only and cannot be cleared from the console. Ship it to your SIEM if you want it retained off the machine.
Nothing is dropped silently
Every drop, mask, aggregation, failover and policy change increments a counter the console renders. If DPLens discards an event, it is because you configured something that discards events, and the number is on the Overview attributed to the rule that did it.
This is a design rule, not a feature: a log agent that loses data quietly is worse than no log agent, because you believe you have coverage you do not have.
One kind of drop is decided by the receiver rather than by you: an OpenTelemetry collector such as ClickStack can refuse individual records, and DPLens drops those rather than re-sending them for ever. Each one is counted on the destination's card, and drops while a cache is draining are written to the audit log. Because a collector — or anything in the network path pretending to be one — can cause such drops, use HTTPS and an ingestion key to your collector, and watch those counters. See Confirmed hand-off.
Hardening checklist
| Leave the console on loopback unless you need it remotely | |
| If you need it remotely, set the allowed networks tightly | Settings |
| Install your own console certificate | |
| Use TLS to every destination that supports it | Choosing a destination |
| Send to ClickStack over HTTPS, with an ingestion key sealed in the secret store | Sending to ClickStack |
| Keep certificate validation on | |
| Configure masking before you configure delivery | Pipeline |
| Set Allowed networks on every syslog and NetFlow receiver | Sources |
| Ship the audit log to your SIEM | Logs and health |
| If you use DPLens Manager, keep enrolment tokens short-lived | Enrolling an agent |
| Keep the default service account rather than LocalSystem | |
| Verify the download before installing | Verifying a download |
| Back up the configuration and licence | Backup and recovery |
Reporting a vulnerability
If you believe you have found a security issue in DPLens, tell us before you tell anyone else, and we will work with you on a fix and a disclosure timetable.
Please include what you found, how to reproduce it, the version affected, and how you would like to be credited. We will acknowledge your report, keep you updated while we investigate, and credit you when we publish the fix unless you ask us not to.
Please do not test against systems you do not own or have permission to test.