This week’s defence AI landscape is defined by a collision between operational urgency and democratic accountability: the US military is racing to embed AI deeper into classified networks and intelligence workflows, while Congress, the EU, and the UK are scrambling to impose governance before capabilities outpace oversight. The Anthropic–NSA–export-control saga has crystallised a new and unresolved tension — that the most powerful AI security tools may simultaneously be the biggest security risks.
Top story: Anthropic’s Mythos AI penetrated nearly all NSA classified systems in hours during an authorised red-team drill, triggering US export controls and a global shutdown of the model — the first time export controls have been applied directly to an AI model rather than hardware.
Mythos Cracked NSA Systems in Hours, Triggering Historic AI Export Controls
TechSpot · Risk
During an authorised red-team exercise, Anthropic’s Mythos AI reportedly penetrated nearly all of the NSA’s classified systems — not in weeks, but hours. The revelation prompted the Trump administration to issue export controls restricting Mythos to US citizens only, the first time export controls have ever been applied directly to an AI model rather than hardware. Unable to verify nationality at scale, Anthropic shut down Mythos globally for all users with 90 minutes’ notice, cutting off allied intelligence services, researchers, and its own foreign-national employees — exposing just how disruptive frontier cyber-AI models have become for national security governance.
EU Publishes Nine-Point AI Cybersecurity Blueprint for Critical Infrastructure
Euronews · Regulation
The European Commission unveiled a structured Action Plan on AI and Cybersecurity on 7 July, setting out nine actions across three goals: making frontier AI safe for cyber defence, hardening EU critical infrastructure, and scaling Europe’s own AI capacity. The plan commits ENISA and the Joint Research Centre to building a secure testing platform for AI in critical sectors by Q4 2026, and a European Blueprint for structured access to advanced AI models — though Euronews noted the EU’s uncomfortable dependency on US firms like Anthropic for the most capable tools. With AI Act enforcement beginning 2 August 2026, the plan signals Europe is moving from lawmaking to operational security posture.
Senator Warren Presses Pentagon and Seven AI Firms Over Secret Military Contracts
Federal News Network · Regulation
Senator Elizabeth Warren wrote to Defence Secretary Pete Hegseth and seven AI companies — including OpenAI, Google, Microsoft, SpaceX, NVIDIA, AWS, and Oracle — demanding the full terms of their classified network agreements be disclosed to Congress by 20 July. Warren warned the deals may give the DoD ‘free rein to engage in domestic surveillance’ or build autonomous weapons without human oversight, and said it is ‘impossible to assess any safeguards’ without seeing the contracts. The pressure follows the Pentagon’s May agreements to deploy these companies’ AI on Impact Level 6 and 7 classified networks, amid ongoing litigation with Anthropic over the same ethical red lines.
UK AI Security Institute Ranked Mythos Top Model for Offensive Cyber
Nextgov/FCW · Strategy
As the NSA lost access to Anthropic’s Mythos 5, reporting confirmed that the UK’s AI Security Institute had independently tested frontier models on a cyber range — and ranked Mythos highest for offensive and defensive cyber capability, ahead of Claude Opus 4.6 and OpenAI’s GPT 5.4. The UK’s NSA counterparts also confirmed access to Mythos through the AI Security Institute, underlining Britain’s active role in evaluating frontier AI for intelligence purposes. The findings reveal a growing transatlantic competition to benchmark — and access — the most capable cyber-AI tools before adversaries do.
US Army Issues RFI for Agentic AI on Classified Military Networks
ExecutiveGov · Strategy
The US Army has issued a Request for Information for agentic AI platforms deployable on its most sensitive classified networks — including JWICS and SIPRNet at Impact Level 6 — enabling intelligence analysts to build, deploy, and oversee AI agents that can autonomously retrieve information and conduct tasks in disconnected or bandwidth-limited combat environments. The procurement signals a major operational shift: moving AI from assistive tool to autonomous agent within the classified intelligence stack. Defence practitioners should note the Army is simultaneously seeking AI autonomy for UAS and next-generation aviation systems under a separate $29 million, five-year programme.
https://www.executivegov.com/articles/army-contract-opportunities-june-2026-ai-mshorad-drones