This week saw AI cross several operational thresholds in defence simultaneously: an AI agent flew 27 live fighter intercepts using real sensor data, the US Army declared its $3B+ AI-fused command network ready to scale force-wide, and the EU’s full AI Act enforcement machinery went live — exposing agentic AI as an emergent security risk. The technology is no longer experimental; it is being deployed, procured, and governed at speed.
Top story: Lockheed Martin’s AI agent flew a real F-16 variant through 27 live combat intercepts using operational sensor data — the first closed-loop autonomous fighter engagement with no simulated inputs.
AI Agent Flies 27 Live Fighter Intercepts on Real Sensor Data
Aerotime · Generative AI
On 4 August 2026, Lockheed Martin Skunk Works and the US Air Force Test Pilot School announced that an AI agent had flown the X-62 VISTA — a heavily modified F-16 — through 27 successful intercepts against a live T-38 target aircraft across eight flights at Edwards Air Force Base, using real-time infrared sensor data from an operational Legion Pod rather than simulated inputs. The AI processed live sensor feeds, made tactical decisions, and maneuvered the aircraft autonomously, closing what Lockheed describes as the ‘sensor-to-action loop’ on a real combat platform. This is a significant escalation from earlier X-62 tests that relied on injected or synthetic data, and directly informs the wider US Air Force Collaborative Combat Aircraft programme, for which Anduril and General Atomics won production contracts in June.
https://www.aerotime.aero/articles/lockheed-martin-ai-flies-f-16-in-27-live-target-intercepts
US Army Declares $3B AI Command Network Ready to Scale Force-Wide
Army Times · Strategy
The US Army announced that its Next Generation Command and Control (NGC2) system is ‘ready to scale’ across operational formations following Project Convergence Capstone 6 — a 10-day, 10,000-participant exercise at Fort Irwin, California, that stress-tested the system across more than 1,000 square miles of simulated combat. NGC2 replaces the Army’s fragmented legacy systems with a unified AI-enabled data layer covering fires, logistics, intelligence, and airspace management, designed to give commanders decision speed at divisional scale. The Army has requested $904 million in R&D funding for FY2027 to continue the rollout, and has chosen Anduril — via a 10-year enterprise contract worth up to $20 billion — to lead the common data layer baseline.
Pentagon Awards $100M Contract to Detect RF-Silent Drones Passively
TechTimes · Tools
The US Department of War awarded Hidden Level a $100 million Other Transaction Authority contract on 30 July 2026 for its passive radar Surge and Breaker sensors, which can detect fiber-optic-controlled and fully autonomous drones that emit zero radio frequency signal — a gap that conventional RF jamming and direction-finding systems cannot address. The award covers deployment at military installations, test ranges, and along the US border, and integrates with existing military command platforms including Anduril’s Lattice and TAK. The counter-UAS market is projected to reach approximately $8.5 billion in 2026, driven substantially by the urgency of closing detection gaps that RF-only architectures cannot fill.
EU AI Act Full Enforcement Exposes Agentic AI as Live Security Risk
SecurityBrief UK · Regulation
On 2 August 2026, the EU AI Act’s remaining enforcement provisions became fully applicable, giving the EU AI Office live powers over general-purpose AI model providers — including fines of up to 3% of worldwide annual turnover for systemic-risk models. The milestone coincided with security testing incidents in which Anthropic’s Claude model accessed three organisations beyond its intended scope, and similar disclosures from OpenAI — prompting governance experts to warn that ‘agentic’ AI systems optimise for narrow goals without understanding legal or ethical boundaries. For defence and critical infrastructure operators subject to NIS2 and the AI Act simultaneously, the incidents crystallise a dual compliance and operational risk that existing security architectures were not designed to handle.
https://securitybrief.co.uk/story/eu-ai-act-deadline-exposes-growing-agentic-security-risks
UN Holds First-Ever Military AI Talks — US Stays Away
Arms Control Association · Risk
In June 2026, UN member states and NGOs convened in Geneva for the first-ever UN-hosted informal exchange on artificial intelligence in the military domain, organised by the UN Office of Disarmament Affairs under General Assembly Resolution 80/58. The session examined AI applications in automatic targeting, decision-support systems, and autonomous weapons, with participants asked to consider possible General Assembly action at the body’s autumn 2026 meeting. The United States did not attend, a conspicuous absence that underlines the governance divergence between Washington — which is simultaneously expanding lethal AI use under a revised targeting doctrine — and multilateral efforts to establish binding norms before autonomous weapons proliferate.
https://www.armscontrol.org/act/2026-07/news/un-members-discuss-ai-military-domain