Economic pressures will drive unchecked AI deployment, outpacing safety science and risking civilization’s control by late 2026, according to ARIA.

The United Kingdom’s Advanced Research and Invention Agency (ARIA), has warned that humanity risks running out of time to address escalating safety threats from rapidly evolving AI systems, as their capabilities surge ahead of control mechanisms.

In an interview with The Guardian, ARIA’s program director, David Dalrymple, had cautioned that economic incentives could compel firms to deploy powerful AI before comprehensive safety science catches up, potentially destabilizing security and economies. Dalrymple stressed the need for immediate mitigation strategies such as monitoring and restrictions, rather than relying on perfect reliability guarantees that may arrive too late.

The ARIA program director forecasts that, by late 2026, AI could automate a full day’s worth of research and development, outcompeting humans in economically vital areas and accelerating self-improvement loops beyond regulatory grasp. Dalrymple highlighted a disconnect between public authorities and AI developers on breakthrough timelines, urging governments not to presume advanced models’ dependability amid commercial pressures. Developing safeguards for AI in essential infrastructure such as power grids remains a priority for his ARIA team.

These concerns align with the UK AI Security Institute’s (AISI) inaugural Frontier AI Trends Report that documents two years of testing on over 30 leading models from 2022 to October 2025.

AI development outpacing human-vigilance ramp-up
The report reveals AI now succeeds at apprentice-level cybersecurity tasks 50% of the time, up from under 10% in early 2024 data, with one model tackling expert tasks needing a decade of human expertise. Also:

  • Autonomous cyber operation durations double every eight months, while software challenges taking experts over an hour saw 40% success rates by mid-2025.
  • In chemistry and biology applications, models surpass PhD experts by up to 60% on open-ended queries, generate feasible lab protocols since late 2024, and outperform humans by 90% in wet-lab troubleshooting.
  • Self-replication benchmark scores jumped from 5% in 2023 to over 60% in 2025, though real-world feasibility lags.
  • Safeguards show uneven progress: Jailbreak-resistance improved 40-fold for some biological misuse cases between models six months apart, yet vulnerabilities persist in every tested system.
  • Agentic scaffolds and reasoning chains drive autonomy gains, narrowing proprietary-open source gaps to 4–8 months. While dual-use potential boosts science and defense, misuse risks in cyber, bio, and persuasion grow.

Dalrymple has described civilization “sleepwalking” into this shift, calling for urgent technical work to comprehend and constrain advanced AI behaviors before they overwhelm oversight.