Weekly signal

This week (Oct 12–20, 2026) the manufacturing sector saw coordinated movement on three practical fronts for agentic AI: (1) an OT safety alert from Booz Allen showing frontier models can traverse multi-vendor plant networks and reach physical equipment in minutes; (2) new commercial pushes to move agents from pilot to production via governed control-planes and partner services (Boomi + Jade Global) aimed specifically at factories and supply chains; and (3) vendor and integrator offerings that package domain-aware agent libraries, multi-agent orchestration, and edge-first runtime choices (TCS Industrial Intelligence Foundry and industry webinars showing small/edge LLMs in pilots) that make shop-floor deployment more tractable.

What changed

1. Offensive-capability test: Booz Allen published lab findings demonstrating that unnamed frontier models, run as autonomous agents, completed all eight OT attack objectives in an isolated multi-vendor manufacturing testbed and moved physical equipment in short order—one run went from perimeter compromise to controller actions in ~16 minutes; the report lists concrete mitigations (segmentation, hardened jump hosts, monitoring of controller writes, backups) for defenders.

2. Governed agent rollouts: Boomi and implementation partner Jade Global announced an expanded partnership to bring a data-first Agent Control Plane, Agent Control Tower, and manufacturing-specific solution patterns to plant-floor and supplier deployments—explicitly positioning governed agent lifecycles as the bridge from PoC to production (announcement dated Oct 12–13, Tokyo / global).

3. Industrial-scale authoring & edge focus: Tata Consultancy Services launched an Industrial Intelligence Foundry that bundles configurable agent libraries, multi-agent orchestration, and governance aimed at product development and factory operations; A3/Automate webinars and manufacturing analysis from the week emphasized edge/physical AI tradeoffs (latency, determinism, RISC‑V / on-device NPUs) and small-model + RAG patterns used in pilots.

What to do with it

• Treat the Booz Allen result as urgent risk input: test segmentation, remove dual-homed systems, restrict pre‑authenticated sessions, and instrument controller/program writes and SCADA project imports now.
• Start small, governed pilots that mirror production constraints: adopt an Agent Control Plane pattern (policy, lifecycle, observability) and prefer phased on‑prem / edge pilots before granting agents write privileges.
• Re-evaluate runtime targets: for safety/latency use cases push inference to edge devices or constrained LLMs with strict tool access; reserve cloud agents for non-safety planning and analysis.
• Operational checklist: map high-consequence assets, adopt agent identity & ownership, require human approval rails for physical actions, keep tested backups and known-good PLC configs.

(Primary sources below.)