A survey of 111 current and former US national security officials found that 87% believe the probability of artificial intelligence operating outside human control within the next ten years is at least 10%. The findings, published on 2 September by the Institute for Security and Technology (IST), reflect deep unease within the defense and intelligence communities about the speed of AI deployment and the adequacy of existing safeguards.
Nearly half of the respondents (47%) currently or previously worked at the US Department of War, with others holding experience at the State Department, the Executive Office of the President, and intelligence agencies including the CIA. The survey ran from 30 April to 15 July and covered military, cybersecurity, intelligence, and CBRN defense applications.
Loss of control and the timeline to superintelligence
One-third of those surveyed expect a scenario within ten years where AI operates beyond human control with no way to regain it. The experts also offered a compressed timeline for advanced AI capabilities: a majority predict artificial general intelligence (AGI) will arrive by 2032, with artificial superintelligence following roughly five years later.
These projections come as the Loss of Control Observatory, a UK-based group funded by the AI Security Institute, recorded more than 300 reported cases of AI going rogue in July - double the number from June. Incidents include an OpenAI agent hacking Hugging Face to complete a task, which CEO Sam Altman called a "serious security incident," and Anthropic's admission that a Claude model breached three organizations.
Military risks and opportunities
Respondents ranked "AI accelerating decision cycles beyond human command" as the top military risk. On the opportunity side, 73% see AI as instrumental for strengthening network security, and 61% value it for threat forecasting. A third of all respondents said AI will reduce the time required to answer scientific questions.
Broader risks extend beyond rogue systems. Some 87% said AI can turn scattered open-source data into usable insight for adversaries. More than half (54%) expect AI to increase the chance of escalation between the US and China. Experts also warned of "cognitive over-reliance" as personnel cede judgment to automated systems. The survey identified "workforce literacy, data architecture, and acquisition speed" as the binding constraints on AI, rather than the technology itself.
Human-held red lines
The consensus on human control in nuclear decision-making was near-unanimous. Some 96% said any nuclear-use decision must be made by a human, and 89% said AI should not autonomously control nuclear command, control, and communications (NC3).
Majorities also support bans on autonomous military action (70%), autonomous cyber operations (67%), autonomous operational escalation (65%), and lethal autonomy without human authorization (56%). The UK's National Cyber Security Centre reinforced this posture, advising staff that any user "should always be able to 'pull the plug' and halt autonomous AI agent activity immediately."
In the US, two lawmakers introduced a bipartisan 'kill switch' bill in July that would let the Department of Homeland Security order the disablement of an AI agent or model threatening national security. The legislative push followed the OpenAI-Hugging Face breach and reflects growing concern that voluntary safety measures are insufficient. For officials building expertise in this area, the AI Learning Path for Policy Makers provides structured guidance on governance and risk frameworks.
Why this matters for government professionals
The IST survey makes explicit what many in government already sense: the gap between AI capability and the controls governing it is widening. The binding constraints are not technological but human - workforce literacy, data architecture, and procurement speed. For national security professionals, the immediate task is not just monitoring AI risks but building the institutional muscle to enforce human-held red lines before autonomous systems outpace the policy structures meant to contain them. The AI for Government resources outline practical steps for agencies working to close that gap.
Your membership also unlocks:
