Superintelligence as a security risk, states between bans and control

The security policy engagement with artificial intelligence has shifted since 2023 from a primarily economic and innovation policy question to a question of arms, non-proliferation and international order. Human rights organisations and the United Nations have been pushing for years for an international law framework for autonomous weapons systems. In August 2024 a UN report explicitly called for a binding treaty by 2026, supported by a joint call from the UN Secretary-General and the International Committee of the Red Cross ¹⁸. The Stockholm International Peace Research Institute SIPRI follows the Geneva negotiations as a specialist observer and provides methodological contributions to the assessment of military AI systems ⁵⁵+145 SIPRI.

In parallel, an institutional landscape developed for the assessment of advanced AI. The United Kingdom established an AI Safety Institute after the Bletchley Park summit, which was initially oriented towards model evaluations and security research ²⁰. The institute now publishes a 'Frontier AI Trends Report' under its new name AI Security Institute, which systematically tracks developments in advanced AI models ³⁶. The US organisation METR, which specialises in technical model evaluation, summarised in May 2026 the common elements of security policies of leading frontier AI providers and thus created a first comparative basis for civil society and regulatory assessment ⁵⁷.

NATO and International Institutions

At alliance level NATO adopted a revised AI strategy on 10 July 2024, which continues the 2021 strategy, names AI-driven disinformation for the first time explicitly as a risk, and requires an alliance-wide landscape for testing, evaluation, verification and validation ¹⁹. The NATO Innovation Agency DIANA launched its annual programme for accelerating military innovation in February 2026, which explicitly also addresses AI applications ⁵⁶. The Atlantic Council released a report in May 2026 in which it argued that NATO must significantly accelerate its AI integration to prevail in coming algorithmic warfare ²⁹.

From an economic policy perspective, the framework until early 2025 was shaped by the Biden administration's export controls on advanced AI chips to China. This line has been reversed since the beginning of Donald Trump's second term, as an analysis by the Bloomsbury Intelligence and Security Institute from February 2026 outlines ¹⁴. Brookings in May 2026 assessed under the heading 'Game Over' a de facto loss of the Chinese AI chip market for US providers, because Chinese buyers are increasingly switching to domestic alternatives ⁴¹.

Export Controls and National Competition

The rule-of-law-oriented US specialist portal Lawfare argued in August 2026 that the 25 percent revenue tax to the US state lacked a clear legal basis and could be challenged in court ⁵⁰. The White House released 'America's AI Action Plan' in July 2025, which consolidates the Trump administration's strategy to secure the United States' leading role in the AI race against China and to roll back regulation ⁵¹. In the EU, a comprehensive horizontal legal framework was created with the AI Regulation, which comes into force in stages and subjects providers of AI models with general-purpose use (GPAI) to special obligations.

China, in turn, according to IAPP at the beginning of 2026, issued a dense series of new regulations in the areas of data protection, generative AI and security assessment and thereby visibly sharpened its own governance framework ⁵⁴.

The debate on existential risks from AI has further broadened in 2026. The German-language specialist service all-about-security.de summarised on 10 September 2026 how the warnings from notable developers fuel the political debate on AI existential risks ⁴⁰. The Atlantic Council already sketched eight ways in January 2026 how AI would shape geopolitics in the current year, including arms competition, standard-setting and institutional fragmentation ³⁰.

At state level, the USA, China, the United Kingdom, the EU, Germany and increasingly India are in focus. In the USA, the Trump administration is driving a relaxation of chip export controls ⁶+214 Bloomsbury Intelligence and Security Institute41 Brookings and has set its own security policy framework with National Security Directive NSPM-11 from 2 June 2026 ²⁶. The White House's 'America's AI Action Plan' from July 2025 brings together the strategy ⁵¹. The Council on Foreign Relations assesses the order as centralisation of federal AI oversight with clear priority for national security ²⁷.

Lawfare also questions the legal admissibility of the new turnover tax on chip exports ⁵⁰. In Congress, representatives of the left wing of the Democrats, including Senator Bernie Sanders and Congressman Greg Casar, introduce a ban on superintelligence ⁷+19 Nextgov/FCW. At the level of US intelligence services, the NSA is expanding the use of AI to process vast datasets according to Defense One ³⁸ whilst according to the libertarian-conservative magazine Reason, the CIA produces AI-generated reports for local police departments ³⁹.

Foreign Policy documented in February 2026 an openly conducted dispute between the Pentagon and the AI company Anthropic over terms of use and military applications ³⁷.

The German Bundestag discussed the implementation of the EU regulation and decided on it on 11 June 2026 ¹⁰+211 Deutscher Bundestag23 Bundesregierung. The parliamentary process is well documented. The Bundesrat dealt with the draft in March 2026 ⁴⁷. The Digital Affairs Committee spoke out on 25 March 2026 in favour of the Federal Network Agency as the central market supervision authority ⁴⁹. It adopted the draft on 26 March 2026 ⁴⁸. The relevant printed matter 21/6404 contains the draft of the AI Market Supervision and Innovation Promotion Act ²⁵.

The Federal Network Agency assumes the central coordination role in the national implementation ²⁴. The AfD group in parallel introduced motions on AI sovereignty ¹². The publicly funded ZDF channel illuminated in July 2026 in a specialist report the risks of military AI use and contextualised the debate for a wider German audience ⁵².

At international level, NATO with its AI strategy ¹⁹, the DIANA Innovation Programme 2026 ⁵⁶ and the Atlantic Council report on algorithmic warfare ²⁹ are central players. So are the United Nations with the debate on autonomous weapons and the forthcoming CCW Review Conference ¹⁷+418 Human Rights Watch32 Human Rights Watch33 UN Office for Disarmament Affairs55 SIPRI as well as the World AI Cooperation Organization (WAICO) headquartered in Shanghai announced by China in July 2026 ⁵. India presents itself as an independent governance actor with the AI Impact Summit and the New Delhi Declaration ³⁴.

The Carnegie Endowment is exploring in July 2026 a parallel US-China approach to AI safety as a possible way out of pure rivalry ³¹. The British security institute, known as the AI Security Institute since February 2025, focuses on misuse risks for chemical and biological weapons, cyber attacks and crime ²⁰+136 UK AI Security Institute. The Stockholm International Peace Research Institute SIPRI laid out a framework in February 2026 for responsible procurement of military AI ⁴⁵ and accompanies the Geneva GGE negotiations as a technical observer ⁵⁵. According to IAPP analysis, China's own governance is intensifying in early 2026 through new rules on data protection, generative AI and security assessment ⁵⁴.

In civil society and academia, the Future of Life Institute with its call for a moratorium ³⁵ and prominent AI researchers such as Geoffrey Hinton and Yoshua Bengio as signatories stand out ⁸+246 TIME58 CyberScoop. The organisation METR specialising in model evaluation provides a comparative overview of security policies of leading frontier AI providers ⁵⁷. From the defence sector, according to research by Correctiv from April 2026, the German companies Helsing, Stark and Rheinmetall are relevant, which develop AI for military applications ¹³.

The market-liberal Handelsblatt portrayed Helsing in March 2026 as a German provider that sees AI-enabled drones as a central component of the future Bundeswehr ⁴². Think tanks including RAND, Brookings, CSIS, Atlantic Council, Carnegie, Council on Foreign Relations, the Bloomsbury Intelligence and Security Institute and CNAS shape the analytical debate. With their 2026 reports 'Red Lines' ⁴⁴ and on US-China competition and military AI ⁵³, they provide assessments on armaments dynamics, export controls and existential risks ¹+132 RAND3 RAND4 Brookings5 Brookings6 Brookings15 CSIS16 CSIS21 RAND27 Council on Foreign Relations29 Atlantic Council30 Atlantic Council31 Carnegie Endowment for International Peace41 Brookings.

The established US foreign policy journal Foreign Affairs describes in an article from 20 May 2026 entitled 'China's AI Arsenal' the Chinese build-up of military AI capacities ²⁸.

In the US Congress, Senator Bernie Sanders and Representative Greg Casar introduced the 'Ban Artificial Superintelligence Act' in early September 2026. The bill aims to prohibit the development and deployment of superintelligent AI and suspend the further development of advanced systems until security rules are established ⁷+19 Nextgov/FCW. The Washington Post reported on 3 September 2026 that the proposal was responding to uncontrolled AI incidents ⁷. TIME described on 8 September 2026 growing pressure on both sides of the Atlantic to ban superintelligent systems ⁸.

Nextgov placed the bill within a broader week of legislative initiatives that also affected AI oversight at the Pentagon ⁹. The scene channel Telegram vx-underground commented on the bill on the same day in deliberately satirical and distorted tone and should be valued as opinion rather than as a primary source ⁴³.

On the executive side, President Trump issued the National Security Presidential Memorandum NSPM-11 on Artificial Intelligence on 2 June 2026 ²⁶. The Council on Foreign Relations assessed the order a few days later as a clear centralisation of federal AI oversight with priority for national security and simultaneously exposed weaknesses in accountability and civilian oversight ²⁷. In parallel, Defense One documented on 8 September 2026 that the NSA wanted to deploy AI systems for systematic analysis of its data holdings ³⁸.

Reason reported on 28 August 2026 that the CIA produced AI-generated situation reports for local police departments ³⁹. Foreign Policy had already described an openly conducted dispute between the Pentagon and Anthropic over usage policies on 25 February 2026 ³⁷. On 11 September 2026, the Telegram channel OSINTdefender reported, citing an Anthropic threat intelligence report, that state-aligned actors from China and Russia had used the company's models for weapons development, surveillance and influence operations. Independent confirmation of the report's contents remains outstanding ²².

On export controls, the Trump administration reversed its previously strict line. Since July 2025, shipments of the Nvidia H20 to China have been permitted again. Since December 2025, the more powerful H200 has also been allowed in exchange for a payment of 25 per cent of the turnover to the US state ⁶+114 Bloomsbury Intelligence and Security Institute. The US think tank Brookings argues in an analysis from 30 May 2026 that the relaxation could be justified provided superintelligence were not imminent ⁶.

In a parallel contribution from 20 May 2026, Brookings concludes, however, that the USA had effectively lost the Chinese AI chip market because Chinese buyers were increasingly turning to domestic alternatives ⁴¹. CSIS places the development in an 'arms race in export controls' and calls for restraint ¹⁵. The CSIS contribution from June 2026, oriented towards transatlantic security policy, argues fundamentally for restraint in AI grand strategy ¹⁶. The rule-of-law oriented US expert portal Lawfare classified the 25 per cent turnover payment in August 2026 as legally challengeable and named possible plaintiffs ⁵⁰. The White House 'America's AI Action Plan' from July 2025 remains the strategic framework for this line ⁵¹.

In the EU, the key obligations of the AI Act for GPAI providers have been in effect since 2 August 2025. The EU Commission's enforcement powers apply from 2 August 2026. Legacy models must comply by 2 August 2027 ¹⁰. In Germany, the legislative path can now be traced without gaps. The Bundesrat dealt with the draft in early March 2026 ⁴⁷. The Bundestag's Digital Committee spoke out for the Bundesnetzagentur as the central market surveillance authority on 25 March 2026 ⁴⁹ and adopted the bill on 26 March 2026 ⁴⁸.

The Bundestag passed the national implementation law on 11 June 2026. The Federal Government confirmed this on the same day ²³. The associated printed document 21/6404 was published on 10 June 2026 ²⁵. The AI Market Surveillance and Innovation Promotion Act names the responsible regulatory authorities and came into force on 29 July 2026 ¹¹. The Bundesnetzagentur took on the central coordinating role in implementation with a press release from 29 July 2026 and acts as the national contact point and market surveillance authority ²⁴. On 12 June 2026, the Bundestag additionally discussed motions from the AfD faction on AI sovereignty ¹².

In the field of autonomous weapons, the deadline until 2026 set by the UN Secretary-General and the ICRC for a binding treaty has passed without formal negotiations being initiated ¹⁷+118 Human Rights Watch. The UN Group of Governmental Experts on lethal autonomous weapons systems convened in Geneva in early September 2026, as documented by the UN Office for Disarmament Affairs ³³. SIPRI attended the GGE meeting as a technical observer ⁵⁵ and had already presented a framework for responsible procurement of military AI in February 2026, which explicitly targets procurement processes in NATO states ⁴⁵.

Human Rights Watch reported on 7 September 2026 that the meeting ended without treaty negotiations, but a growing majority of states were demanding formal negotiations ³². The final decision will be made at the CCW Review Conference in November 2026 ¹⁷. According to a single investigation by Correctiv from April 2026, the Bundeswehr concluded procurement contracts in early 2026 with Helsing, Stark (26 February 2026) and Rheinmetall (22 April 2026) for loitering munitions with AI-assisted target recognition.

The systems are to be delivered to the brigade in Lithuania from 2027 onwards. Independent confirmation of these details is outstanding ¹³. The market-liberal Handelsblatt portrayed Helsing in March 2026 as a German provider positioning drones with AI-assisted target recognition as the future of the Bundeswehr ⁴². The public broadcasting ZDF channel illuminated the opportunities and risks of military AI use in armed conflicts in a specialist report in July 2026 ⁵².

At the international level, China announced the founding of the WAICO at the World Conference on Artificial Intelligence in Shanghai in July 2026. Brookings places this as a visible expression of a growing divide between the USA and China in AI governance ⁵. Foreign Affairs describes in the article 'China's AI Arsenal' from 20 May 2026 a systematic Chinese build-up of militarily usable AI ²⁸. IAPP reports in parallel about a dense sequence of new Chinese regulations on data, generative AI and security assessment at the beginning of 2026 ⁵⁴.

RAND analysed the perspective of the People's Liberation Army on AI integration as key to the 'intelligentisation' of the Chinese armed forces in March 2026 ²¹ and analysed the AGI race as a distinct international security problem in May 2026 ¹. In August 2026, RAND described new risks of AI development for US security in relation to China ². A RAND analysis from February 2026 warned of a security gap between AI and weapons of mass destruction ³.

Brookings assessed existential risks from AI in April 2026 and discussed possible responses ⁴. The Carnegie Endowment proposed in July 2026 a parallel US-China path to AI security that was intended to enable cooperative security standards despite systemic rivalry ³¹. The Center for a New American Security submitted a report on US-China competition and military AI in April 2026, followed by the report 'Red Lines' from 12 June 2026, which names red lines in US policy towards Chinese leading-edge AI ⁴⁴+153 CNAS.

India hosted the AI Impact Summit 2026 in February. The Foreign Ministry confirmed on 20 February 2026 the adoption of the 'New Delhi Declaration' on Artificial Intelligence at the conclusion of the summit ³⁴. The Atlantic Council outlined eight ways in which AI would shape geopolitics in 2026 on 10 January 2026 ³⁰ and submitted a report on 1 May 2026 stating that NATO would need to accelerate its AI integration to prevail in future algorithmic warfare ²⁹. The NATO Innovation Platform DIANA launched its annual programme in February 2026, which includes military AI applications ⁵⁶.

The British AI Security Institute has oriented its work towards national security since its renaming in February 2025, particularly towards misuse for chemical and biological weapons, cyberattacks and crime ²⁰. In June 2026, it published a 'Frontier AI Trends Report' on developments in high-capability models ³⁶. The NATO strategy from July 2024 remains the reference framework for alliance-wide military AI use ¹⁹. The organisation METR, specialised in model evaluation, published a comparative overview of the safety policies of leading frontier AI providers in May 2026 and describes common elements and gaps ⁵⁷.

On the civil society side, the Future of Life Institute called for a ban on the development of superintelligent AI in October 2025 with more than 700 signatories, including Geoffrey Hinton and Yoshua Bengio, until scientific consensus on safe control and broad public consent exist ⁸+135 Future of Life Institute. TIME documented the call on 22 October 2025 ⁴⁶. The US expert portal CyberScoop confirmed the broad support on 23 October 2025 and placed the moratorium as an expression of growing mainstream concern ⁵⁸.

The German specialist service all-about-security.de concluded on 10 September 2026 that the warnings from developers had significantly sharpened the political debate on AI existential risks ⁴⁰.

The security policy AI debate runs along several parallel lines of conflict. First, the question of superintelligence regulation is on the table, in which a ban proposal by Sanders and Casar ⁷+19 Nextgov/FCW competes with positions that justify loosening export controls by arguing that superintelligence is not immediately to be expected ⁶. Both positions refer to the same facts but draw opposing conclusions from them. Warnings from model developers and the expert public ³⁵+340 all-about-security.de46 TIME58 CyberScoop additionally shift the debate framework towards existential risks. The overview of company-owned security policies provided by METR ⁵⁷ simultaneously shows how strongly self-regulation by providers still determines the field.

European and American Regulatory Approaches

Second, a divergence is emerging between a rules-based European approach with the AI Regulation ¹⁰+711 Deutscher Bundestag23 Bundesregierung24 Bundesnetzagentur25 Deutscher Bundestag (dserver)47 Deutscher Bundestag hib48 Deutscher Bundestag hib49 Deutscher Bundestag and a more geopolitically shaped US line that combines export controls, state participation and executive security management ⁶+514 Bloomsbury Intelligence and Security Institute26 The White House27 Council on Foreign Relations41 Brookings51 The White House. The announcement of the Chinese WAICO ⁵, the parallel intensified Chinese domestic regulation ⁵⁴, the Foreign Affairs analysis on the Chinese AI arsenal ²⁸ and the RAND study on military AI integration by the People's Liberation Army ²¹ point to an institutional consolidation of this rift.

The Brookings assessment from May 2026 ⁴¹ and the Lawfare classification of the turnover tax ⁵⁰ suggest that US chip policy is under economic and legal pressure. CNAS simultaneously sets a reference framework with "Red Lines" ⁴⁴ and the analysis of US-China military AI competition ⁵³ that recommends tough security policy responses. The Carnegie proposal to establish parallel US-China security standards ³¹ marks a counter-draft whose political viability remains open. India's AI Impact Summit and the New Delhi Declaration ³⁴ show that further governance poles are forming.

International Law and Autonomous Systems

Third, the international legal framework for autonomous weapons systems remains open. The expiry of the 2026 deadline without the start of negotiations ¹⁷+118 Human Rights Watch, the result of the Geneva expert meeting in September 2026 with a clear majority for negotiations but without a formal resolution ³²+233 UN Office for Disarmament Affairs55 SIPRI and the parallel national procurements as Correctiv describes for Germany ¹³ and as Helsing publicly classifies in the Handelsblatt ⁴² point to a discrepancy between normative aspiration and operational practice. SIPRI addresses this gap with a framework for responsible procurement of military AI ⁴⁵ that offers practical entry points for defence ministries. The reliability of Correctiv's information on contract partners and delivery timescales remains unclear due to the lack of independent confirmation.

Fourth, government use of AI in the internal and external security area of the USA is visibly expanding. NSA data analysis ³⁸, CIA reports for local police forces ³⁹ and the dispute between the Pentagon and Anthropic ³⁷ show an unresolved framework between operational use, fundamental rights and the terms of use of providers. The Anthropic threat intelligence report reported via OSINTdefender ²² also points to state misuse attempts with commercial models, which provides arguments both for stricter export and access rules and against further loosening. At the alliance level, NATO attempts to structure military AI integration through DIANA ⁵⁶ and the 2024 strategy ¹⁹; the Atlantic Council simultaneously urges greater speed ²⁹.

The CCW review conference in November 2026 ¹⁷ is a fixed point at which it will become apparent whether a multilateral regulation is still politically viable. Whether the Sanders-Casar draft finds majorities in Congress remains open. The sources available do not provide reliable statements on this.