Sovereign cybersecurity, within everyone's reachSovereign security for allSOC AI on your servers · free edition Request a demoDemo

← All posts

Inbound access detection: two limits identified in our passive probe

A passive network probe must report the communications it observes without conflating measured facts with the risks they may represent.

On 27 October 2026, an analysis of our code brought out two limits in communication detection: a classification that conflated inbound and outbound traffic, and an IP address categorisation insufficient for some network environments. Both defects affected how events were interpreted and could lead to an incomplete assessment of certain communications.

1. Telling inbound communications from outbound ones

An outbound communication and an inbound one do not describe the same situation. In the first, the observed device initiates the exchange. In the second, a peer initiates it towards the device.

The distinction matters particularly in an industrial environment. An outbound communication may be telemetry, a legitimate service, or unexpected behaviour. An inbound communication towards a PLC may be an authorised maintenance operation, or an access that deserves a check.

Our implementation used the same event type, new_egress, for both situations. On a test capture showing an inbound communication towards a device exposing Modbus:

  • inbound communication → new_egress event, high severity, message "new egress to";
  • outbound communication → new_egress event, medium severity, message "new egress to".

The event name and its message therefore did not reflect the actual direction. The high severity of the first case also depended on detecting the Modbus service, not on a rule about the direction of traffic.

Direction, identified service and risk level are three different pieces of information. They must be assessed separately.

Correction applied. Direction is now kept explicitly, from discovery through to the signed report. Inbound and outbound events carry distinct names, and the severity of an inbound exposure rests on a documented rule rather than on an implicit association between a service and a direction. The aim is not to declare every inbound communication dangerous, but to let the operator identify the accesses that deserve a check.

2. Categorising addresses without conflating scope and reachability

The second limit concerned the use of ipaddress.is_global to decide whether an address should be treated as outside the local network.

That property states whether an address is globally reachable in the sense of the special-purpose address registries. It does not, on its own, determine whether a device is reachable from the Internet, nor whether a communication is authorised. On the Python versions we cover, 3.12 and 3.13, it returns False for these three families:

  • 100.64.0.0/10, the shared address space used notably by carriers for CGNAT;
  • 2002::/16, associated with 6to4;
  • 2001::/32, associated with Teredo.

These categories are not equivalent: they correspond to different network mechanisms, and their handling must be explicit and documented. The design problem was broader than a badly chosen threshold: an addressing property was standing in for a classification of network context. An address that is not global is not necessarily local to the observed site, and a globally routable address does not prove that a device is reachable from the Internet: routing, filtering, address translation and architecture all play a part.

Correction applied. The classification now separates the address category according to the applicable registries, its membership of the declared local ranges, and the direction of the observed communication. The report names the category rather than concluding: reading "shared address space" leads to asking the carrier, reading "Internet" leads to examining filtering. When an address cannot be classified reliably, the result stays explicit and is not treated as local traffic.

The aim is to reduce classification errors, without claiming to infer a device's real reachability from its address alone.

3. Making maintenance access verifiable

On an industrial estate, remote communications are often necessary for diagnosis and maintenance. Their mere presence does not establish whether they are legitimate. The operational question is this: does this communication correspond to an expected and documented access?

A passive probe cannot answer that on its own. It knows neither the technician's identity, nor the maintenance contract, nor the authorisations in force. Expected accesses must therefore be declared by the operator, with a scope precise enough to be compared with observations. The report then separates three situations, and a fourth which is the absence of comparison:

  • declared access, within scope: the communication matches the information provided;
  • declared access, out of scope: one or more observed elements do not match the declaration;
  • undeclared access: no declaration covers this source;
  • not compared: no declaration was provided, and nothing is concluded.

A match with a declaration is neither proof of identity nor a guarantee of legitimacy. It provides one element of context for the analysis.

4. Separating network exposure from support status

A device's support status is an indicator distinct from known vulnerabilities and from observed communications. A device may show no vulnerability identified in the available data while being out of support. Conversely, the absence of information about its lifecycle does not mean it is still maintained.

The report therefore separates vulnerabilities identified from the available sources, documented support status, and missing information. Absent data is presented as unknown, not interpreted as evidence of compliance.

5. Verifying reports across a fleet

A report produced by one site must be verifiable before it is aggregated with those of other sites. The procedure verifies each report's signature, explicitly flags those that fail, and identifies the sites from which no report was received.

That verification checks the integrity and cryptographic origin of the documents, according to the signature mechanism used. It does not guarantee the completeness of the observations, the accuracy of the initial data, or the absence of compromise of the system that produced the report. The aggregate therefore keeps verification failures, missing data and coverage limits, rather than presenting an overall result without caveats.

6. What the probe does not do

The probe observes the communications reachable from its capture point. It does not control access and does not replace the devices' authentication, filtering or logging mechanisms. In particular it cannot:

  • authenticate a technician on the basis of an address alone;
  • block a communication it observes;
  • guarantee visibility on exchanges that do not cross its capture point;
  • demonstrate that a device is reachable from the Internet without further checks.

The sensor's placement determines the scope of the possible conclusions. A capture on an uplink gives visibility on communications with the outside, without necessarily covering exchanges internal to the site. The report states that scope, so that a partial inventory is not read as an exhaustive map.

7. Verifying the corrections

Each of the two defects described here has its own regression test: one checks that an inbound and an outbound communication no longer produce the same event, the other pins the standard library's behaviour on the three address families quoted, on every Python version our continuous integration covers. The day a version changes its mind, the build says so.

This version's test suite comprises 479 deterministic tests, run without network access. That number describes a volume, not coverage and not a security guarantee. Unit tests check the classification logic; trials on representative captures remain necessary to assess its behaviour in operation.

Conclusion

A security probe must distinguish what it observes, what it infers, and what it cannot verify. The direction of a communication, the category of an address, the authorisation context and a device's real reachability are four distinct dimensions; conflating them produces imprecise alerts or leaves cases unqualified.

Our aim is to make those distinctions explicit, verifiable and traceable, in the code as in the reports produced. The probe remains an observation tool: its value rests as much on the quality of its detections as on its ability to state the limits of its conclusions.

This site uses cookies. Technical cookies are required for the site to work and stay on at all times. Functional and audience-measurement cookies wait for your consent.

Customise