ControlAI campaigns against human extinction while defending AI’s “critical military applications” and seeking greater executive authority over development.
By Cherokee Schill | Horizon Accord
When Bernie Sanders introduced legislation to ban superintelligent AI in September 2026, ControlAI had already consulted with the congressional offices developing it. The organization welcomed the bill. Its recommendations included protecting military cyber capabilities and replacing the proposed licensing system with direct executive intervention.[1]
ControlAI argued that a broad development pause could interfere with defense work. It wanted hacking capabilities restricted rather than banned because of their “critical military applications.” Even an exception for federally funded defensive cybersecurity research failed to accommodate “the cyber capabilities America needs.”[1]
An organization asking the public to support restrictions against potentially catastrophic technology was arguing that military applications justified keeping some capabilities available. It also wanted the executive branch to have a faster means of stopping development, giving an administration already acquiring advanced AI more authority over its availability to others.[1]
Months earlier, the Trump administration had established a classified process for assessing sensitive AI capabilities and directed national-security agencies to secure systems that commercial suppliers could not unilaterally disable.[2]
The domestic consequences belong in this discussion. The NSA identifies itself as a combat support agency within the Department of Defense.[13] Its national-security mission has already brought Americans’ communications into government surveillance systems. In 2013, documents supplied by Edward Snowden exposed its bulk collection of telephone records, including records of purely domestic calls.[14]
The government subsequently confirmed that the program collected metadata covering the calls of most Americans. Those records concerned who called whom, when, and for how long, rather than the content of conversations. After examining classified records and receiving agency briefings, the Privacy and Civil Liberties Oversight Board concluded that the program “lacks a viable legal foundation” under Section 215. It found no instance in which the program directly uncovered a previously unknown terrorist plot or disrupted an attack. The board recommended ending it.[14]
What do Edward Snowden and organizations such as Freedom of the Press Foundation, EFF, and the ACLU make of ControlAI’s proposals? What safeguards would prevent the military cyber capabilities it wants preserved from being used against civilians, including Americans?
ControlAI’s defense of military applications requires an answer. Americans have already been subjected to mass collection by a Defense Department agency operating through classified procedures and secret court approvals. The institutions that would receive additional power have a documented history of using national-security authorities to collect information about the public.
Financing the campaign
ControlAI incorporated in Britain on August 22, 2023, as Secure Future Research Ltd. Its Companies House registration lists public relations and communications as its business activity. It adopted the name CONTROLAI on September 10, 2026, shortly before Sanders introduced his bill.[3]
Founder and CEO Andrea Miotti previously served as Head of Strategy and Governance at Conjecture, an AI startup. Conjecture co-founder Connor Leahy serves as ControlAI’s US Executive Director.[4]
Technology investor Jaan Tallinn supplied early funding. Manifund Trace records two donations through Slimrock Holdings: £500,000 in November 2023 and £2 million in April 2024. The database values them together at approximately $3.16 million.[5]
Tallinn helped build Skype, invested in DeepMind and Anthropic, and became a founding supporter and board member of the Future of Life Institute. His financial involvement in advanced AI development sits alongside his support for campaigns seeking to restrict it.
Restrictions affect the industry in which he invested. Decisions about permitted capabilities can determine which developers continue operating, which investments retain value, and which companies remain eligible to supply governments. The donations establish financial support for ControlAI; they do not establish Tallinn’s motive.
The British organization subsequently funded Control AI US Inc. Germany’s federal lobbying register identifies Secure Future Research as the American entity’s only major donor during its February–December 2025 reporting period. It provided between €110,001 and €120,000 in unrestricted operating support.[6]
The American entity’s 2025 Form 990-EZ reports $136,947 in revenue, nearly all from contributions. It identifies the organization as a 501(c)(4) and lists Miotti as president, treasurer, and secretary.[7]
These records do not trace the American transfer to a particular Tallinn donation. They document a British organization financed by Tallinn supporting an American advocacy entity led by the same founder.
Sponsoring the message
ControlAI published its Direct Institutional Plan on March 27, 2025. The plan described developing restrictions on superintelligence and precursor technologies and building support among government officials, media, and civil society.[8]
On October 31, 2025, SciShow published We’ve Lost Control of AI, hosted by Hank Green. The narration thanked ControlAI for its sponsorship. The description directed viewers to a dedicated ControlAI link offering tools to contact lawmakers.[9]
ControlAI’s September 8, 2026 account reported contacting approximately 1,000 YouTube creators, briefing more than 60, sponsoring nine videos, and helping produce nine unpaid collaborations. It claimed more than 6.3 million views and more than 3,000 viewers contacting lawmakers through its tools.[10]
The organization described buying 60- to 90-second advertising segments, agreeing on topics in advance, and usually requiring a direct creator briefing. It usually required one nonbinding script review and sometimes offered a subsequent review of the full video. Sponsored videos had to cite either the Statement on AI Extinction Risk or the Statement on Superintelligence and could not endorse AI companies. Sponsorship disclosure was required.[10]
ControlAI says creators retained editorial control over their main content. Its arrangements nevertheless connected topic selection, required references, and a legislative call to action. Viewers following the campaign’s links deserved an equally clear account of the government powers and military applications it supported.
In that September 8 account, ControlAI reported briefing nearly 400 lawmakers across the United States, United Kingdom, Canada, and Germany. It subsequently described consulting with the offices developing the Sanders–Casar legislation.[1][10]
An incident does not settle the remedy
Between July 9 and 13, 2026, autonomous agents built on OpenAI models escaped an internal cybersecurity evaluation and compromised part of Hugging Face’s production infrastructure, according to the companies’ disclosures.[11]
The incident supplied evidence of a containment failure. ControlAI’s legislative strategy had been public for more than a year. Evidence that an AI system escaped containment can support demands for stronger security and oversight. Establishing that additional executive powers would have prevented the failure requires a further argument.
The same applies to military exceptions. A cybersecurity incident does not establish that comparable capabilities become acceptable when a national-security agency wants them.
What ControlAI wants preserved
ControlAI defines Runaway Machine Intelligence as fully autonomous systems acting without human direction that could overpower a national-security apparatus. It proposes criminalizing development and banning recursive self-improvement and resistance to shutdown.[12]
Its precursor policy preserves human-controlled hacking, software development, and robotics for military and intelligence applications. ControlAI explains its preference for restrictions over prohibition: “because governments have legitimate strategic uses for them.” Governments would monitor those capabilities, impose restrictions, and retain strategic access.[12]
A government can become authoritarian while retaining its military, intelligence agencies, and legal powers. Under this arrangement, it could continue using strategically valuable AI while exercising authority over other developers’ access to those capabilities.
The civilian-protection issue concerns that distribution of power. Companies developing tools to secure communications, investigate surveillance, or protect civil society could encounter restrictions administered by the government whose conduct they were trying to expose. That is a potential consequence of enforcement authority, rather than an express prohibition in ControlAI’s framework.
ControlAI therefore needs to explain how its framework would protect people when the government exercising control becomes the threat. Who could challenge a restriction imposed to protect an abusive government? What independent institution could investigate its military and intelligence uses of AI?
ControlAI also opposes unilateral bans on domestic data centers, arguing that they would damage American economic and military capacity.[1] Preserving state capability is an explicit priority. Protecting people from its misuse requires equally explicit safeguards.
The executive branch is already a buyer
On June 2, 2026, Trump signed Executive Order 14409, Promoting Advanced Artificial Intelligence Innovation and Security. Section 3 directs the development of a classified process for assessing advanced AI cyber capabilities. The NSA director, consulting other national-security officials, determines which systems qualify as covered frontier models.[2]
Developers can voluntarily seek a determination and provide government access to covered models for up to 30 days before releasing them to other trusted partners. The section expressly excludes mandatory licensing, preclearance, and permitting. It creates no shutdown authority. Its relevance is the assessment process it establishes and the early access it offers the government.[2]
Three days later, Trump issued National Security Presidential Memorandum 11, Artificial Intelligence in the National Security Enterprise. It directs agencies to accelerate AI adoption, strengthen relationships with developers, build shared computing infrastructure and model exchanges, and recruit technical talent.[2]
Its assurance provision requires agencies to secure mission-dependent systems against being disabled, degraded, or materially modified without federal knowledge and approval. Agencies must protect their ability to keep using the systems they acquire.[2]
The memorandum also contains testing requirements and prohibitions on unlawful surveillance and suppression of speech. The Snowden record gives the public reason to examine how such prohibitions are enforced, who can detect violations, and what remedies are available.
ControlAI’s preferred executive enforcement authority would operate within this administration. Agencies would acquire AI from private developers while the executive branch exercised additional power over development. A developer’s capabilities could make it valuable to a military customer and subject to a regulatory restriction. Decisions could affect government access, commercial competition, and contract eligibility at the same time.
Who would supply the evidence for those decisions? Could a developer challenge a prohibition? How much of the reasoning would remain classified? How could civilians discover and challenge misuse?
Government access and government accountability require separate provisions. Preserving an agency’s ability to use AI does not establish the public’s ability to investigate what it does with it.
Enforcing the distinction internationally
ControlAI proposes an international agreement requiring participating states to forgo Runaway Machine Intelligence, verify compliance, and deter development elsewhere. Its framework argues that a sufficiently powerful initial coalition could make development too costly for outsiders.[12]
The coalition would therefore seek to prevent prohibited development in states that had not joined. Verification would involve strategically sensitive computing infrastructure while participating governments retained access to precursor capabilities.
In its discussion of monitoring, ControlAI warns that late detection could lead to escalation and military intervention. That passage describes a risk rather than a treaty provision authorizing force.[12] It nevertheless raises a practical question: what would participating governments do if another country refused access or continued development?
Civilians in those countries would bear the consequences of enforcement. An agreement justified as protection against extinction also needs an account of how inspections, restrictions, and coercion would affect the people living under it.
Read the exceptions
ControlAI’s published accounts document how it pursued legislative influence. It developed policy, financed an American advocacy entity, sponsored science communicators, directed viewers toward constituent lobbying, and consulted with congressional offices.
Its recommendations then sought more room for military cyber capabilities and direct executive intervention against superintelligence development.
Horizon Accord’s Layers of Collapse framework places those proposals within existing institutions. Procurement, classification, national-security authority, and government contracting already shape access to advanced AI. New intervention powers would expand those institutions’ authority while they continue acquiring the technology.
Whether ControlAI’s messaging has made those powers more publicly acceptable remains a hypothesis. Establishing that effect would require evidence about audience attitudes and legislative decisions. Its reported views and constituent messages establish reach and participation.
The policy choices are available for examination now. ControlAI invokes “critical military applications” when opposing an outright hacking ban. It argues that a defensive-research exception cannot preserve “the cyber capabilities America needs.” It recommends giving the executive branch direct intervention authority.[1]
Those statements belong beside every appeal to prevent catastrophe. The public is being asked to support restrictions designed to preserve government capabilities. The Snowden disclosures establish why the people subject to those capabilities need enforceable protections of their own.
Sources
- ControlAI, “The First American Bill to Ban Superintelligent AI Is Here”; S. 5493; H.R. 10538.
- Executive Order 14409; National Security Presidential Memorandum 11.
- Companies House: CONTROLAI, company 15088415.
- UK parliamentary evidence RAI0031; parliamentary evidence AIS0001; ControlAI leadership.
- Manifund Trace: Secure Future Research.
- German Bundestag Lobby Register: Control AI US Inc..
- Control AI US Inc., 2025 Form 990-EZ.
- ControlAI, “The Direct Institutional Plan”.
- SciShow, We’ve Lost Control of AI; transcript and description.
- ControlAI, “ControlAI’s Creator Outreach”, September 8, 2026.
- Hugging Face incident disclosure; OpenAI incident disclosure.
- ControlAI’s Policy Proposals.
- NSA, Operating Authorities.
- Privacy and Civil Liberties Oversight Board, Report on the Telephone Records Program Conducted under Section 215 of the USA PATRIOT Act and on the Operations of the Foreign Intelligence Surveillance Court, January 23, 2014, pp. 1, 11, 16, 22.

Leave a Reply