The research leaves a simple division implicit, so here it is. Great Britain has about 6.8 GW of storage across roughly 1,400 units [5], an average near 4.9 MW per unit [1]. Four hundred of those average about 1.94 GW [2], which lands just past the 1.8 GW Narezzi calls the shock absorber [14]. The 400-unit figure is a secured-loss ceiling divided by average unit size, and the ceiling is the operative number, not a share-of-fleet finding: GB is engineered to ride through the sudden loss of 1,320 MW inside statutory frequency limits, with a higher allowance for rarer events, past which the obligation is only to hold near 49.2 Hz for up to 60 seconds [13]. Below that, the first stage of low frequency demand disconnection at 48.8 Hz sheds blocks of customers automatically, with relays doing the work [14][15].
Then the much smaller number in the same document. Afzal and colleagues found in 2025 that a load-altering attack using 15 percent of a fleet's power could push frequency outside normal operating bounds [7], from which Centrii estimates 11 to 21 compromised 2 MW units could destabilise a regional grid [8]. That is 22 to 42 MW [3], between 36 and 68 times smaller than the 1,500 one-megawatt ERCOT units cited elsewhere in the same work [4]. Help Net Security notes one figure describes a regional test network and the other a whole interconnection, and that the two are never reconciled [9]. For a defender, the unresolved gap is the finding: one figure implies a campaign against a fleet, the other a campaign against three sites.
The attribution problem follows from how the money works. Optimisers trade a battery's output across power markets on the owner's behalf, so the credential that books revenue is the credential that moves megawatts [18]. Narezzi named KrakenFlex as an example of a cloud-based optimiser where large capacity can sit under one cloud base [19], and describes the target as the smallest set of control planes whose combined fast-acting MW exceeds the regional secured-loss ceiling [16]. Centrii puts entry through such a platform at 35 to 70 percent success over two to five weeks at intermediate skill [6]. Whatever distinguishes a hostile dispatch from a revenue dispatch lives in that platform's authentication and command records rather than on the asset owner's screens, and all the asset owner holds is a contract.
The 92.1 percent headline probability carries the least weight of anything here. It is the output of 10,000 Monte Carlo runs seeded with assumed compromise rates, an assumed 15 percent annual gain in attacker capability, and 25 to 35 percent annual storage growth, against a scenario that has not yet occurred [11]. The drop to 61.4 percent under IEC 62443 certification and quarterly drills [10] is a 33 percent relative reduction [5] produced by the model's own assumption about how well those controls work.
Narezzi says he has a candidate signature for telling the two swings apart [21]; the published interview stops before he describes it. Until one is specified and tested against live dispatch data, detection for a BESS operator sits in someone else's cloud logs and attribution sits in a commercial agreement.
