AI in the Trading Room: ADEX Energy, Virtual Power Plants, and the Adversarial Robustness Gap

Booth C5.416, Hall C5, EM-Power Europe. ADEX Energy — the Sofia-based SaaS vendor whose one-line pitch reads "Modular SaaS for energy transition — AI forecasting, trade automation & BESS optimization for RES operators" — was showing something the show floor is, quietly, full of: a stack of AI agents with real market access. Forecast module reads weather and generation history. Trading module reads the intraday order book and submits bids. Optimization module reads state-of-charge from a portfolio of BESS assets and schedules dispatch against the trading module's plan. Each module runs on its own cadence. All three touch money.

The class of platform

ADEX is representative, not unique. Next Kraftwerke, Entelios, Flexpower, Sympower, In.power and half a dozen EM-Power Europe exhibitors ship variations of the same architecture: a portfolio of distributed energy resources, aggregated into a virtual power plant, dispatched through a trading strategy that is increasingly implemented as an ML model or an LLM-assisted decision loop.

This is the operational core of the energy transition. It is also, from a security standpoint, one of the least-scrutinized parts of the stack.

What we discussed at the booth

The failure modes that keep VPP operators up at night are not the ones the AI-safety literature emphasizes. They are:

• Adversarial inputs on the price feed. If a market data source is spoofed or delayed, the trading agent submits bids priced against a fictitious book. Imbalance settlement charges do the rest.
• Model drift under regime change. A forecasting model trained on 2022–2024 weather and dispatch patterns is measurably worse in 2026's higher-penetration, higher-volatility market. Silent degradation is expensive.
• Cross-module trust assumptions. The trading module assumes the forecast is honest. The optimization module assumes the trade schedule is honest. Neither module verifies the other. Compromise one and you compromise the chain.
• Operator copilot injection. Every VPP platform is adding a natural-language interface ("why did we sell 40 MWh at 03:00?"). That copilot reads DSO messages, trader chat and market notifications — all attacker-influenceable.

The partnership shape

TrendGuru's role in a VPP-operator conversation is complementary, not competitive with the domain expertise. Concretely:

• Agent registry and blast-radius review. Every deployed agent catalogued with its market access, its credentials and its worst-case loss. Reviewed quarterly, updated on every strategy change.
• Input-integrity monitoring. Signed, timestamped price and weather feeds; anomaly detection on the distributions the models actually consume.
• Adversarial evaluation. Structured red-team runs against the forecasting and trading models, with the results tracked release-over-release and fed back into retraining.
• Copilot hardening. Indirect-injection test suites against the operator-facing NLP layer, plus egress controls on the runtime.

The regulatory backdrop

The relevant instruments are stacking. NIS2 designates energy operators as essential entities with 24-hour incident notification. The EU AI Act's Article 15 requires demonstrable robustness and cybersecurity for high-risk AI systems, and market-facing dispatch models sit squarely inside that scope. REMIT II tightens the surveillance regime on wholesale energy markets, including for algorithmic trading. Together, they turn the security posture of a VPP's AI stack from a nice-to-have into a compliance line item.

A virtual power plant is a supply chain of models. Sign every link, monitor every output, and assume every input is hostile.