This week’s defence AI news is defined by a collision between ambition and accountability: Western militaries are racing to embed AI across every domain — from the UK’s RAID taskforce to the US Marines going AI-first — while Google DeepMind’s landmark security roadmap and the IDF’s SMASH Hopper procurement expose how inadequately the industry has grappled with the risks of deploying autonomous agents and AI-assisted weapons at scale. The gap between deployment speed and governance frameworks has never been wider.
Top story: Google DeepMind published a landmark AI Control Roadmap treating its own agentic AI as a potential insider threat — a defence-in-depth model that directly challenges the assumption that alignment training alone is sufficient for military-grade AI deployment.
Google DeepMind Treats Its Own AI Agents as Insider Threats
Fortune · Risk
Google DeepMind published its AI Control Roadmap on 18 June, outlining a zero-trust, defence-in-depth framework for agentic AI that assumes alignment training may fail and builds structural containment on top of it. The framework borrows directly from insider-threat cybersecurity models, applying real-time monitoring and dynamic access controls to AI agents that could act faster and at greater scale than any rogue human employee. For defence and security practitioners deploying agentic AI on classified networks, this is the most detailed public blueprint yet for governing systems that cannot be fully trusted — and a signal that even the labs building frontier AI do not believe alignment alone is sufficient.
UK’s RAID Taskforce Faces Procurement Scepticism From Defence Founders
Sifted · Strategy
Prime Minister Keir Starmer launched the Rapid AI Delivery (RAID) taskforce during London Tech Week, designed to accelerate deployment of AI-enabled tools across the British military and help personnel make faster decisions while keeping pace with adversaries. But UK defence tech founders and investors are warning the initiative risks being undermined by the Ministry of Defence’s notoriously slow procurement processes, drawing an unflattering contrast with the US, where industrial agility is treated as a strategic asset. The story matters because European defence startups raised €2.3bn in 2026 — over 100% up year-on-year — and whether RAID can actually channel that energy into the Armed Forces is the defining test for UK defence AI credibility.
https://sifted.eu/articles/uk-defence-raid-rapid-ai-in-defence-taskforce
US Marines Reveal Five-Initiative Plan to Become AI-First Force
Federal News Network · Strategy
The US Marine Corps has detailed five operational initiatives transforming it into an AI-first warfighting force, following a secretarial directive that fundamentally redefines the Corps’ posture — requiring workflows, tactics, and operational concepts to be reimagined as if born in the age of AI. The plan empowers the Pentagon’s Chief Data and AI Office to cut through procurement bureaucracy, with data-sharing denials now requiring senior-level justification within seven days. Defence practitioners should note this is no longer a strategy document but an operational mandate with teeth, and marks the most concrete branch-level AI transformation blueprint to emerge from any US military service this year.
GCHQ Warns of ‘Narrowing Window’ as Russia Escalates Hybrid Cyber Attacks
Intelligent CIO Europe · Risk
GCHQ Director Anne Keast-Butler warned at Bletchley Park that the UK and its allies face a narrowing window to maintain technological advantages as Russia escalates daily hybrid activity — including cyberattacks and infrastructure targeting — and the NCSC now handles approximately four nationally significant cyberattacks against Britain every week. The proposed AI-enabled national cyber defence system could take up to five years to fully deploy, a timeline that industry experts have sharply criticised as incompatible with the tempo of current attacks. For security professionals, the story highlights a structural governance gap: the UK is publicly committing to machine-speed AI cyber defence while simultaneously acknowledging it will not be operational for half a decade.