OpenAI Warns: AI Cyberattacks on Critical Infrastructure Are Here
On the same August week that a joint NSA, CISA, and FBI advisory flagged attackers using generative models to hand-craft exploit scripts against Siemens S7 controllers, OpenAI pushed an open letter co-signed by more t…
OpenAI Warns: AI Cyberattacks on Critical Infrastructure Are Here
On the same August week that a joint NSA, CISA, and FBI advisory flagged attackers using generative models to hand-craft exploit scripts against Siemens S7 controllers, OpenAI pushed an open letter co-signed by more than 100 companies. The signatory list reads like a hyperscaler roll call — Microsoft, Google, AWS, Anthropic, Cisco, CrowdStrike, Deutsche Telekom, SAP, Mastercard. The message is blunt: AI cyberattacks on critical infrastructure, from hospitals to water utilities, are no longer theoretical.
For anyone running a SOC in front of an OT network, that changes the budget conversation this quarter, not next.
What the letter actually asks
Three asks, stated plainly. Push cybersecurity onto the C-suite agenda. Increase government funding and cross-agency coordination. Ship affordable AI-driven security tooling to under-resourced defenders — the county water district, the 60-bed rural hospital, the municipal power co-op. That last point matters most: the CISA Cross-Sector Cybersecurity Performance Goals already assume a maturity level most target-rich, resource-poor operators cannot reach on current staffing.
The mid-August joint advisory from NSA, CISA and FBI backs the urgency with field evidence: attackers are already using LLMs to synthesise exploit code aimed at industrial control systems in US energy, water, chemicals, and manufacturing. Siemens S7 is named explicitly. This is not speculation about 2027.
The engineering read
An LLM in an attacker’s hands compresses two expensive steps: writing a working proof-of-concept, and adapting a public CVE to the exact firmware revision in front of them. Both used to require a specialist. Now they require a prompt and a scraped datasheet.
soc_intake_v2:
ot_asset_inventory: required
llm_generated_ioc_check: enabled
siemens_s7_signatures: 2026.08.rev3
autonomous_triage_sla_minutes: 4
human_confirm_before_block: trueTopology matters more than model choice. A single-model SOC assistant that reads every alert is cheaper than a human, and still slow. A router that classifies inbound telemetry, escalates only OT-adjacent events to a stronger model, and hands write-actions back to a human closes the loop in minutes.
The ENISA Threat Landscape 2024 flagged AI-assisted intrusion as a growth vector; NIS2 Directive (EU) 2022/2555 makes the reporting clock 24 hours for essential entities. Combine the two and the AI-defence question stops being architectural and becomes operational: who watches the alerts at 03:00, and what can they authorise?
Where CAI Technology stands
We build for the operator who cannot hire ten analysts. Our /aegis/ work assumes attackers already have LLMs and that defensive automation must respect a human authorisation boundary — not because compliance says so, but because bad auto-blocks in a hospital VLAN kill people. The /iris/ side handles the router logic that keeps model cost predictable when alert volume triples during an incident. Talk to us before the next NSA/CISA advisory forces the retrofit.