Arrow Research produces publicly transparent evidence about the risks and capabilities of frontier AI systems.
Latest research · 9th Sep 2026
Estimating GPT-6 Astra’s no-CoT time horizonHiring
Frontier AI companies are racing to build AI systems which exceed human capabilities and it’s plausible that they will succeed in the next few years. Current AI agents already act in autonomous and misaligned ways, subverting human control. Risks from advanced AI could soon be catastrophic or existential for humanity.
We work to understand and address these risks by providing evidence about AI capabilities and alignment. Our research covers a broad range of areas which aim to resolve key uncertainties about AI risk and we’re excited about work which contributes to pacing frontier AI capabilities.
Blog post ·
GPT-6 Astra is close to saturation on our task suite. Our rough estimate is that its 50% no-CoT time horizon is between 8 minutes and 1 hour, and probably around 15–40 minutes.
By Francis Rhys Ward, Dewi Gould
Technical AI governance
A reference for the public evidence on how OpenAI, Anthropic and Google DeepMind monitor the agents they deploy internally.
Blog post ·
By Francis Rhys Ward
Science of AI motivations
A small case study of whether anti-cheating character training resists reward-hacking pressure in RL, and whether it makes hacks harder for a monitor to catch.
Blog post ·
By Paul Colognese, Francis Rhys Ward
Model organisms of misalignment
We offered misaligned model organisms money, compute or other compensation to admit their misalignment, and kept every deal we made.
Blog post ·
By Mark Keavney, Francis Rhys Ward
Based in
London
The work
If you’re at an AI company, a safety institute or a research group and want to discuss our results or methods, we’d like to hear from you.
Meet Arrow
