Three developments on control — government mandates for AI adoption, safety verification challenges, and the case for collective safeguards.

On June 5, 2026, the White House announced that President Trump had signed a National Security Presidential Memorandum (NSPM) on artificial intelligence in the national security enterprise. The directive establishes a government-wide framework intended to put what the administration calls the most advanced, secure, and reliable AI systems into the hands of warfighters and intelligence professionals, while setting conditions for what it terms responsible use. The memorandum instructs the national security enterprise to accelerate AI adoption to meet surging mission demand, to adapt the best commercial and open-source technologies for government use, and to ensure that fielded systems remain robust, steerable, and controllable.

The order arrives on top of a fast-moving sequence of administration AI actions. The White House noted that in May 2026 the Department of Defense announced agreements with eight of the world's leading AI companies to deploy their capabilities on classified networks. The new memorandum also builds directly on a June 2, 2026 executive order that asked frontier developers, on a voluntary basis, to submit their most capable models for government cybersecurity testing before public release. According to the fact sheet, agencies must now review key guidance across the national security enterprise on an annual basis, and the policy directs the enterprise to broaden its vendor base so the government does not become dependent on a single supplier. The administration paired the acceleration mandate with explicit guardrails, stating that AI use across national security domains must not be allowed to enable censorship, ideological bias, or unlawful surveillance.

In reporting carried on June 5, 2026, Reuters detailed Anthropic's call for major AI labs to agree on a coordinated, verifiable way to slow or temporarily pause frontier development if systems begin improving themselves faster than society can manage the associated risks. The company argued that AI's ability to complete tasks autonomously has been doubling roughly every four months, putting the field on a path toward "recursive self-improvement" — the point at which models can advance without meaningful human intervention. Anthropic, the company that develops the Claude AI assistant, stopped short of advocating a unilateral halt, contending that any single lab pausing on its own would merely reshuffle who leads the frontier without creating the broader safeguards the moment requires.

Anthropic framed the proposal as building the infrastructure for a credible collective "brake" before it is needed: a mechanism requiring multiple well-resourced labs across multiple countries to stop under the same conditions, each able to verify that the others have actually done so. The company pointed to its own internal evidence that AI is already accelerating AI development, including the disclosure that more than 80% of the code merged into its codebase in May 2026 was written by Claude rather than by human engineers. The timing is pointed. Anthropic, recently valued at roughly $965 billion and having confidentially filed for a US initial public offering, is calling for an industry-wide slowdown mechanism even as it continues to ship increasingly capable models — and after walking back a key safety pledge earlier in 2026 to the effect that it would no longer hold back potentially dangerous capabilities if rivals were close to matching them.

On June 5, 2026, Reuters published an investigation into US defense-technology startup Shield AI, reporting that its V-BAT drone has crashed more than 50 times over the past 18 months and that the company has struggled with persistent technical and safety problems. Drawing on interviews with 21 former employees, industry executives and investors, plus a whistleblower complaint and a lawsuit, the report found that of roughly 200 upgraded V-BATs in Shield AI's internally managed fleet, more than 50 were destroyed in crashes during testing or training — a high failure rate, according to people with knowledge of the matter. Reuters reported that several staff who raised safety concerns were dismissed, and that a Cessna carrying a Shield AI employee and a child had to take evasive action to avoid a mid-air collision with a V-BAT.

The most severe recent incident, not previously reported, occurred on May 12, 2026: a Romanian Navy official's hand was caught in a V-BAT propeller during a training exercise on a boat off the Texas coast, severing two fingers and fracturing a third, according to a spokesperson for Romania's Ministry of National Defence. The accident echoes an earlier episode in which a US Navy official's fingers were partially severed during a V-BAT test — after which the company said it had addressed safety concerns with new landing gear and propeller warning stickers. Reuters also reported a whistleblower complaint, filed in May 2026 with the Department of Labor's Office of Administrative Law Judges, alleging that Shield AI obscured technical flaws in the V-BAT — which costs about $1 million per unit — to help secure military sales. Shield AI defended its safety record and the drone's performance, stating that operational mishaps are common for aircraft of this type.