Centripetal Limited (Claimant) v. Palo Alto Networks, Inc. (Defendant) UPC_CFI_660/2024 and UPC_CFI_134/2025
Order of 19 December2025[i]
The Mannheim Local Division (LD) of the Unified Patent Court (UPC) has decided in favour of the Defendant, Palo Alto Networks, Inc., in infringement proceedings and a counterclaim for revocation concerning Centripetal Limited’s patent EP 3 652 914. The LD ultimately revoked the patent in its entirety in Germany and France, resulting in dismissal of the infringement claim.
The defendant, Palo Alto Networks, is a US-based global provider of cybersecurity products, software and services. Centripetal sued for direct and indirect infringement of the German and French designations of EP 3 652 914 based on the defendant’s marketing and supply of its cybersecurity platforms in those territories.
The patent addressed challenges in network security arising from the increasing volume and rate of generation of network threat event logs. While network devices can detect and log large numbers of potential threat events, only a small proportion are ultimately reportable to a relevant authority (e.g., an authority responsible for ensuring the security of the network or a regulatory authority). The patent identified a risk that the volume and creation rate of such event logs can overwhelm human cyber analysts, generate backlogs in event queues, and potentially result in reportable events being overlooked.
The patent described an accelerated cyber analysis workflow in which each event log was assigned a ‘reportability likelihood’ indicating the likelihood that the event is reportable to an authority. That likelihood was produced using the combined outputs of two different algorithms – a static algorithm and a machine‑learned algorithm. Event logs were sorted according to this likelihood and transmitted to an analysis system, allowing events with a higher probability of being reportable to be prioritised for investigation.
The interpretation of the terms static algorithm and machine-learned algorithm was central to the LD’s validity assessment.
The LD interpreted a static algorithm broadly as any algorithm that does not change its assumptions automatically, for example, via training. It rejected the claimant's argument that a static algorithm must be limited to human-designed, explicitly programmed algorithms.
A machine-learned algorithm was interpreted as any algorithm capable of changing its assumptions automatically through training. The LD held that the claimed imposed no restrictions on the type of machine learning technique used, the sophistication of the learning process, or the nature of training data. This was contrary to the claimant's arguments that the machine-learning system must learn to take into account 'human bias' or to develop a cyber analyst’s 'gut feeling', which the LD found appeared only in the description and could not be read into the claims.
Applying its claim construction, the LD found that the asserted claim combination lacked novelty over US 2015/0213358 A1 (HAWK).
HAWK disclosed a static algorithm in the form of pre defined Boolean rule sets and score rules used to assign scores to event occurrence items. These rules were fixed and did not learn or adapt automatically, satisfying the ‘static algorithm’ requirement.
HAWK also disclosed a machine-learned algorithm, in the form of a Naïve Bayes algorithm, which the LD characterised as a well-known supervised machine-learning algorithm. The fact that the algorithm relied on pre-defined parameters, statistics and training data provided by humans did not negate its machine‑learning character, as it still adapted based on training. In particular, the LD noted that such an algorithm could learn attributes as ‘good’ (not associated with a potential threat event) or ‘bad’ (associated with a potential threat event), and that a pre-defined feature might shift from one category to another over the course of training.
On this basis, the LD concluded that HAWK directly and unambiguously disclosed the claimed combination of a static algorithm and a machine learned algorithm, and therefore anticipated the asserted claims.
With the asserted claim combination being held to lack novelty, the claimant applied to amend the patent via numerous auxiliary requests. The amendments sought, among other things, to:
- characterise the static algorithm as human designed or heuristic;
- specify particular machine learning techniques (e.g., artificial neural networks (ANNs)); or
- limit how the machine learned algorithm was trained or which features it evaluated.
The LD held that the amended claims of some requests still lacked novelty over HAWK (e.g., those specifying the static algorithm as being human-designed). For the requests where novelty could be acknowledged, the LD held that the amended claims lacked an inventive step in view of HAWK combined with the common general knowledge or further prior art.
For inventive step, the LD applied the approach emerging from recent UPC case law, according to which a claimed solution is obvious, if, starting from a suitable starting point in the prior art, the skilled person would be motivated (i.e., have an incentive) to consider the solution of the invention and implement it as a next step. The starting point need not be the closest prior art; rather, the decisive point is whether the starting point is one that the skilled person would take into account if confronted with the problem to be solvedii .
Further, the use of a particular means may be obvious even without a corresponding specific motivation, if, by its nature, said means, as general means to be considered for a plurality of applications, belongs to the general knowledge of the relevant skilled person, the use of the functionality in question is objectively appropriate in the context to be assessed, and no special circumstances can be identified that would render its application impossible, difficult or otherwise impractical from a technical point of viewiii .
HAWK was considered a suitable starting point because it addressed the same problem as the patent, namely providing improved systems and methods for review and analysis of potential threat events across networks to efficiently manage large amounts of data. The use of ANNs for assessing potential threats in a computer network was held to be part of the common general knowledge before the priority date. The LD therefore held that replacing the Naïve Bayes algorithm in HAWK with a commonly known ANN was objectively appropriate and obvious, absent any teaching or technical difficulty that would make it imperative to use a Naïve Bayes classifier.
Even if ANNs were not part of the skilled person’s common general knowledge in the specific context, the LD found that the skilled person consulting HAWK would have been motivated to find an alternative or improved machine-learned algorithms. A further prior art document (EP 3 018 879 A1, HLCC17) disclosed neural-network-based malware detection and would have led the skilled person to the relevant feature without an inventive step.
Where auxiliary requests relied on more specific technical features – such as retraining based on previously analysed events, or the use of particular domain name characteristics or entropy values – the LD these to be obvious design options within the common general knowledge or in light of further cited prior art.
As a result, none of the claim amendment requests were deemed to be novel and inventive.
The LD’s reasoning can be contrasted with the more structured approach to inventive step set out in other recent UPC decisions, in which the court identifies an objective technical problem, selects one or more realistic starting points, and asks whether the skilled person would – rather than merely could –have arrived at the claimed solution, supported by a pointer or motivation and a reasonable expectation of success.
In the present case, the LD adopted a more streamlined analysis. It did not dwell on the formulation of an objective problem, nor did it explicitly address whether the skilled person would have had a reasonable expectation of success starting from HAWK. Instead, the LD emphasised that no specific motivation may be required where the claimed feature represents a well known alternative within the skilled person’s general technical knowledge and there is no technical barrier to its implementation. This underscores the risk for patentees that broad claims relying on generic AI/ML terminology may be treated as routine design choices unless tied to specific technical constraints.
The LD left open:
- whether a total of 32 auxiliary requests for amendment was reasonable and admissible under R. 30.1 (c) RoP; and
- whether late-filed validity attacks first put forward in the Reply to the Statement of defence to the Counterclaim were admissible under R. 263 RoP.
These issues did not need to be decided because the patent failed on substantive grounds
However, the LD held that prior art documents introduced for the first time in the Defence to the Application to amend the patent were admissible at least insofar as they were used to defend against the claim amendments.
The LD revoked the patent in its entirety in Germany and France. The asserted claim combination of the main request lacked novelty over HAWK, and all amendment attempts were rejected for lack of novelty or inventive step.
Due to the revocation of the patent, the infringement claim was dismissed.
An appeal may be lodged at the UPC CoA within two months of notification of the decision.
[i] https://www.unifiedpatentcourt.org/en/node/160107
[ii] The LD cited UPC_CFI_311/2023mn. 57 and UPC_CFI_14/2023mn. 8.6 and the CoA decisions of 25 November 2025, Amgen v Sanofi-Aventisand Meril v Edwards, as reported here.
[iii] The LD cited Local Mannheim, decision of 2 April 2025, UPC_CFI_359/2023mn. 121, following BGH, decision of 15 June 2021 – X ZR 58/19, GRUR 2021,1277 mn. 47 – Führungsschienenanordnung