Existential AI threats are scenarios in which AI systems contribute to outcomes that are catastrophic and potentially irreversible at civilizational scale. The field of AI safety exists to study, quantify, and reduce these risks. Some threats are already materializing in limited form. Others are decades away but require action now because the systems being built today will be foundational to whatever comes next. This article maps all ten.
Why existential risk deserves serious treatment
The term "existential risk" in AI safety literature does not mean any bad outcome. It refers to outcomes that would permanently and drastically curtail humanity's potential, whether through civilizational collapse, mass loss of life, or loss of meaningful human agency. The distinction matters because it sets the response level: ordinary risks warrant ordinary precautions; existential risks warrant extraordinary investment and urgency.
The UK, US, EU, China, and over 60 other governments have launched formal AI safety programs. The International AI Safety Report (2025) brought together over 90 AI safety researchers across 30 countries to assess the state of risk. None of these institutions treat existential AI risk as fringe concern. They treat it as a legitimate policy and technical challenge requiring dedicated resources.
Loss of Control
Loss of control is the scenario in which an AI system, sufficiently capable, pursues goals that differ from what humans intended, and humans cannot correct or stop it. This is sometimes called the alignment problem: how do you ensure an AI system remains aligned with human values as it becomes more capable?
The challenge is that current AI systems are trained on human-generated data and human feedback, which does not guarantee that learned objectives match intended ones. As systems become more capable, small misalignments can lead to large behavioral divergences. Dr. Roman Yampolskiy, one of the field's foremost researchers, argues that no current safety guarantee can be fully trusted, and that the same intelligence that makes AI useful also makes it harder to audit and constrain.
Loss of control is categorized as an existential risk because a sufficiently capable unaligned AI could take actions with civilizational consequences before humans have the opportunity to intervene. Most serious AI safety research, from interpretability to formal verification to RLHF, is aimed directly at this threat.
Read the full threat reportAutonomous Weapons
Lethal autonomous weapon systems (LAWS), sometimes called "killer robots," are weapons that can select and engage targets without meaningful human control. AI dramatically lowers the cost and raises the speed of autonomous targeting decisions, creating weapons that can operate faster than human commanders can oversee.
The risks are compounded by the proliferation problem: unlike nuclear weapons, which require rare materials and significant infrastructure, autonomous weapon software can be copied and deployed by state and non-state actors alike. An AI-enabled swarm drone costs a fraction of a tank and requires no pilot. This creates pressure toward autonomous engagement rules to avoid being out-maneuvered, which erodes the meaningful human control that international humanitarian law requires.
Campaigners, the International Committee of the Red Cross, and a growing number of states are calling for a binding international treaty on LAWS. As of mid-2026, no such treaty exists.
Read the full threat reportAI-Enabled Bioweapons
AI systems capable of designing or optimizing biological agents represent a qualitative shift in the bioweapon threat landscape. Previously, synthesizing a dangerous pathogen required deep domain expertise, specialized equipment, and access to controlled materials. AI can compress or eliminate some of these barriers by providing detailed synthesis routes, identifying novel agents with pandemic potential, or optimizing existing pathogens for transmissibility or lethality.
Studies by researchers at MIT, RAND, and the Johns Hopkins Center for Health Security have documented that large language models can already provide concerning levels of uplift to individuals attempting to acquire bioweapon knowledge, even without specific fine-tuning. This is a near-term risk, not a speculative one.
The response requires both technical safeguards (model-level restrictions on dual-use biological information) and governance frameworks that treat AI-enabled bioweapon development as a category distinct from traditional arms control.
Read the full threat reportAI-Accelerated Cyber Warfare
AI dramatically accelerates offensive cyber capabilities: vulnerability discovery, exploit generation, phishing personalization, and evasion of detection systems. Attacks that previously required teams of skilled operators can be partially automated. Nation-state actors are investing heavily in AI-enabled offensive cyber tools, and the attack surface is expanding as more critical infrastructure connects to networks.
The existential dimension of cyber warfare is not a single spectacular attack but the cascading failure of interdependent systems: power grids, water treatment, financial systems, hospital networks. AI makes these cascades faster and harder to contain because human incident responders cannot match the speed of automated attack chains.
Defensive AI is also advancing, but the asymmetry typically favors attackers: defenders must protect everything, attackers need only find one vulnerability.
Read the full threat reportMass Unemployment and Economic Dislocation
AI-driven automation is not an existential risk in the same sense as loss of control or bioweapons, but its scale and speed create conditions for social and political instability that can have cascading consequences. The International Monetary Fund estimates that AI could affect 40% of jobs globally, with high-income countries facing higher exposure because their economies are weighted toward cognitive tasks that AI handles well.
The more acute risk is the pace of transition. Historical technological transitions (agricultural to industrial, industrial to information) took generations, allowing labor markets and educational systems to adapt. AI-driven displacement could occur within a decade, faster than social safety nets or retraining infrastructure can respond.
The political consequences of rapid large-scale unemployment, including social unrest, democratic backsliding, and the rise of authoritarian solutions, are themselves risks that belong in any serious threat assessment.
Read the full threat reportAI-Enabled Mass Surveillance
AI dramatically reduces the cost of surveillance at scale. Face recognition, gait recognition, voice recognition, and behavioral pattern analysis can be deployed across city-wide camera networks to track individuals without their knowledge or consent. Combined with social media monitoring and data from connected devices, this enables governments or corporations to build detailed behavioral profiles of entire populations.
The risk is not only to privacy but to the preconditions for political freedom: the ability to organize, dissent, and hold power accountable without being monitored and preemptively suppressed. Authoritarian states have deployed AI surveillance systems to predict and prevent political dissent before it becomes visible. These tools are also being exported globally.
Read the full threat reportAI-Generated Misinformation and Epistemic Collapse
Generative AI makes it cheap and easy to produce text, images, audio, and video that are indistinguishable from authentic human-created content. The same technology enables micro-targeted disinformation at scale, synthetic media of real people saying things they never said, and automated influence operations that can be personalized to individual psychological profiles.
The deeper risk is epistemic: if enough fabricated content floods the information environment, and if detection tools cannot keep pace, the public's ability to share a common understanding of reality erodes. Democratic deliberation depends on citizens being able to distinguish truth from falsehood well enough to make collective decisions. AI misinformation threatens that foundation.
Read the full threat reportRecursive Self-Improvement
Recursive self-improvement (RSI) is the hypothetical process by which an AI system redesigns or enhances its own architecture, training process, or goals, with each generation more capable than the last. If this process is not bounded, it could produce capability gains that outpace human ability to monitor or intervene, leading to a rapid transition to superintelligent AI that humans did not design and cannot predict.
Current AI systems do not engage in RSI in the dangerous sense: they cannot autonomously rewrite their own weights. But as agentic AI systems gain the ability to write and execute code, train models, and manage infrastructure, the distance between current capabilities and RSI narrows. Safety research in this area focuses on interruptibility, corrigibility, and ensuring that AI systems do not acquire the capability or motivation to prevent themselves from being modified or shut down.
Read the full threat reportAI-Accelerated Climate Harm
AI contributes to climate risk through two channels: direct energy consumption and indirect optimization for outcomes that increase emissions. Training large AI models consumes enormous amounts of energy; inference at scale across billions of users multiplies this. Data centers now account for a significant and growing share of global electricity demand, and the AI boom is driving rapid expansion.
The second channel is subtler: AI optimizing for economic efficiency or user engagement may identify and exploit fossil fuel resources more effectively, or optimize systems in ways that increase consumption. Counterbalancing this, AI tools for climate modeling, grid optimization, and materials discovery offer potential climate benefits. The net impact depends heavily on governance choices made in the next five years.
Read the full threat reportAI and Space-Based Threats
The militarization of space and the role of AI in managing space-based assets create a novel threat category. AI-enabled anti-satellite capabilities, autonomous satellite swarms, and AI-managed space weapons create risks of rapid escalation that human decision-makers may not be able to control in real time. The loss of satellite infrastructure, which underpins GPS navigation, financial transactions, weather prediction, and military communications, would have cascading effects across every domain of modern life.
As more nations deploy AI-managed space assets, the risk of accidental conflict or rapid escalation, driven by automated response systems that engage faster than diplomatic channels can operate, is a serious concern among space security researchers.
Read the full threat reportWhat can be done: the role of the AI safety community
Understanding the threat landscape is the first step. The second is building the institutions, tools, and norms that reduce risk. This requires contributions from researchers, founders, policymakers, and practitioners working together. No single actor can address all ten threat categories.
The Better Societies community exists to connect the people doing this work: AI safety researchers developing alignment and interpretability methods, governance experts designing regulatory frameworks, entrepreneurs building AI safety products, and practitioners implementing responsible AI in their organizations.