PST-QEdge
An edge security layer that watches charger traffic on site, isolates the affected charger and protects links with post-quantum cryptography.
Charge point operators, SOC teams and grid operators
AEGIS, EU-funded research for O-CEI Pilot 7
Separate. This works at the charger or in the lab, and GridSEAL works from the back-end.
What it does
- Scores traffic in a container on the charger gateway, so detection doesn’t depend on a trip to the cloud.
- Isolates only the attacked charger through a response playbook, so neighbouring chargers can keep charging.
- Tunnels the gateway-to-back-end link with hybrid post-quantum cryptography, using ML-KEM and ML-DSA. The charger-to-gateway hop keeps its existing protection.
- Learns across sites with federated learning. Raw telemetry stays on site, and only encrypted model updates leave.
- Scores each site’s updates through a gate that quarantines poisoned ones.
- Writes each automated response step to a hash-chained ledger, for use in an incident report.
How it’s deployed
Operators keep their chargers, management system and data. Safety layers are switched on one at a time, and live isolation comes last.
Stage 1
Detection only
The edge model scores traffic and takes no action
Stage 2
Shadow learning
Learning rounds run, and candidate models aren’t deployed
Stage 3
Heartbeat watch
The stall threshold is set at three vendor heartbeats
Stage 4
Isolation dry run
Playbooks rehearse on a sandbox station
Stage 5
Enforced learning
The gate promotes models
Stage 6
Live isolation
Needs the site owner’s signature, and starts on one spare charger
Results so far
AEGIS was tested with Pilot 7 on a live 5G-connected charging network, and that pilot is complete. The figures we publish come from the emulated testbed run of 23 to 30 September 2026. The figures below come from the earlier run in Deliverable D3.
Measured on an emulated Pilot 7 testbed, 23 to 30 September 2026. These aren’t results from a customer site, and they aren’t GridSEAL figures.
| Measure | Target | Result |
|---|---|---|
| Detection accuracy on unseen data | Above 90% | 98.14%, with 1.41% to 1.95% false alarms per node |
| Reaction time at the charger, P95 | Below 20 ms | 12.85 to 13.69 ms per node |
| Post-quantum handshakes | Hybrid on each link | 520 handshakes, P95 4.9 ms, no failures in 7 days |
| Emulated chargers online during attacks | No disruption | 100% over 7 days, longest heartbeat gap 16 s against a 45 s limit |
| Targeted response | Isolate only the attacked charger | 15 playbook runs, and neighbours kept charging |
| Footprint | Up to 15% CPU and 1 GB | About 0.04 cores and 505 MB |
Under poisoned updates, standard federated averaging fell below 30% accuracy, and the AEGIS gate stayed above 90%.
Limits today
- Pilot 7 testing on a live 5G-connected charging network is complete. It hasn’t run at a customer site.
- A deployment needs site access, a mirror port and an isolated test network.
- The gate’s protection against poisoning gets stronger with more sites, and it hasn’t run across five or more.
- It supports compliance work and doesn’t grant compliance. It adds no radio component.
Talk to us about a pilot
We’re looking for sites and test teams to help validate this work. AEGIS components can be released as open source. Ask us which licence applies to your use.
Also from AEGIS
PST-GridTrace
Inventory of devices, firmware and certificates
See the details
PST-ThreatClone
Attack simulation in a sandbox
See the details
the European Union
AEGIS received funding from O-CEI (Horizon Europe, Grant Agreement 101189589), open call Challenge P7C3. This reflects only PinakashieldTech OÜ’s view. Neither the O-CEI Consortium nor the European Commission is responsible for any use that may be made of the information it contains.