The Safety Horizon

The State of AI Safety 2026 — an annual assessment of how fast the frontier is moving, and whether our ability to assure it is keeping up.

Summary

Capability has become infrastructure. Assurance is still a craft. Frontier AI systems now write production code, negotiate workflows, operate as agents across the open internet, and sit inside hospitals, banks, and ministries. The systems crossed from demonstration to dependency faster than any general-purpose technology on record, and faster than our tools for evaluating, interpreting, and controlling them have matured. That mismatch, not any single model, is the central safety fact of 2026.

The flagship volume of the Frontier Series takes stock of artificial intelligence safety at mid-decade. Its central finding is a widening capability–assurance gap: the pace at which frontier systems gain autonomy and reach has outrun the pace at which we can evaluate, interpret, and control them. The report inventories what has gone right, maps where the gaps are widening, sketches four trajectories to 2030, and closes with an eight-point agenda that assigns concrete work to every stakeholder with a hand on the wheel.

Key findings

1. The capability–assurance gap is the master risk Every specific concern this series examines (misuse, systemic fragility, labor shock, epistemic erosion) grows in the space between what systems can do and what anyone can verify about them. Closing that gap is the organizing goal of the safety field.
2. Real progress happened, and it is worth defending Frontier safety frameworks, national AI safety and security institutes, third-party evaluation ecosystems, and interpretability breakthroughs moved safety from aspiration toward engineering discipline. The scaffolding of a mature assurance regime now exists in outline.
3. Agentic autonomy is outrunning the test bench Evaluations built for chat-era models (static prompts, scored answers) cannot certify systems that plan, use tools, spend money, and act over hours. Control and monitoring of agents is the least mature assurance domain in our index, and the fastest-growing source of exposure.
4. The gap is now a deployment problem, not just a lab problem Most incidents in the coming years will originate downstream, in the thousands of businesses, agencies, and startups wiring capable models into consequential processes without the safety engineering the labs apply upstream.
5. Verification is the missing institution Aviation has flight recorders and independent crash investigators; finance has auditors and stress tests. AI still largely asks the world to take developers' word. Independent, adversarial, well-resourced verification is the highest-leverage institutional investment of the decade.
6. The window is open, and it is a window Nothing in our assessment says a safe transition is out of reach. Everything in it says the choices that determine which 2030 we get, Managed Ascent or Hard Landing, are being made now, mostly by people who do not think of themselves as making them.

How to cite this report

The Safety Horizon: The State of AI Safety 2026 — an annual assessment of how fast the frontier is moving, and whether our ability to assure it is keeping up. FAIR Labs Frontier Series, Vol. 01. Fair Artificial Intelligence Research Labs, 2026. Available at https://fairlabs.ai/research/frontier-vol01-the-safety-horizon

Topics: AI safety, frontier models, assurance, evaluations, alignment, governance, safety maturity

FAIR Labs (Fair Artificial Intelligence Research Labs) is a nonpartisan 501(c)(3) research institute in Washington, DC. Its reports may be reproduced with attribution for non-commercial purposes. For interviews, briefings or data requests, use the contact form.

Related FAIR Labs reports