{"id":9,"date":"2026-04-18T22:25:12","date_gmt":"2026-04-18T22:25:12","guid":{"rendered":"https:\/\/eric-clay.com\/?p=9"},"modified":"2026-04-19T00:44:44","modified_gmt":"2026-04-19T00:44:44","slug":"ai-progress-in-four-charts-a-q2-2026-brief-for-policy-makers-and-staff","status":"publish","type":"post","link":"https:\/\/eric-clay.com\/?p=9","title":{"rendered":"AI Progress in Four Charts: A Q2 2026 Brief for Policy Makers and Staff"},"content":{"rendered":"\n

AI progress has been extraordinarily rapid over the past three years. Language models have moved from failing high-school math problems to contributing to open Erd\u0151s problems in ways that leading mathematicians have publicly praised.[1]<\/a><\/sup> This brief covers four trends policy makers should understand.<\/p>\n\n\n\n

1. AI Progress is Accelerating Exponentially in Programming<\/h2>\n\n\n\n
\"METR
METR Time Horizon 1.1: the length of task (in human hours) that frontier AI agents can complete with 50% reliability, 2019\u20132026.<\/figcaption><\/figure>\n\n\n\n

Anthropic, OpenAI, and Google DeepMind are all focused on automating programming in an effort to automate AI research itself. Scaling compute, combined with a technique known as reinforcement learning from verifiable rewards (RLVR), has yielded continual gains in domains where answers can be checked against ground truth: software development, mathematics, and cybersecurity.<\/p>\n\n\n\n

METR\u2019s January 2026 update estimates that the time horizon for LLM coding is doubling roughly every 131 days since 2023, a sharp acceleration from the 2019\u20132025 stitched trend of seven months, and the 2024-onward trend is faster still at 89 days.[2]<\/a><\/sup> If this trend holds, by the end of 2026 we will likely have AI agents capable of multi-week programming and cybersecurity time horizons. The rate of doubling may itself increase as AI is used to build the next generation of AI, producing a disorienting and unpredictable speedup.<\/p>\n\n\n\n

\u201cWe might be 6 to 12 months away from models doing all of what software engineers do end-to-end. I have engineers within Anthropic who say I don\u2019t write any code anymore.\u201d<\/p>\u2014 Dario Amodei, CEO of Anthropic, World Economic Forum, January 2026[8]<\/a><\/sup><\/cite><\/blockquote>\n\n\n\n

2. Offensive Cyber Capabilities Are Growing Particularly Fast as Coding Improves<\/h2>\n\n\n\n
\"UK
UK AISI evaluation: average attack-chain steps completed by frontier models on \u201cThe Last Ones\u201d multi-stage cyber range, plotted against cumulative tokens spent. Mythos Preview (red) substantially outperforms prior frontier models.<\/figcaption><\/figure>\n\n\n\n

The recent announcement of Anthropic Mythos has shaken the industry.[3]<\/a><\/sup> The UK AISI study on Mythos shows that AI is now able to autonomously conduct sophisticated end-to-end cyberattacks against weakly defended systems.[4]<\/a><\/sup> As model capability improves and agent harnesses become more sophisticated, AI can take increasingly complicated actions in the real world across compressed time horizons. The primary bottleneck to agentic AI has been lack of reliability on complex multi-step tasks. In software development and cybersecurity, that bottleneck is rapidly dissolving.<\/p>\n\n\n\n

\u201cThe current systems are getting pretty good at cyber. You need to make sure that the defences are stronger than the offences.\u201d<\/p>\u2014 Demis Hassabis, CEO of Google DeepMind, India AI Impact Summit, February 2026[9]<\/a><\/sup><\/cite><\/blockquote>\n\n\n\n

3. Open-Weight Models Lag Frontier by 6 to 12 Months<\/h2>\n\n\n\n
\"Epoch
Epoch Capabilities Index (ECI), closed-weight (teal) vs. open-weight (pink) models. The leading open-weight models consistently trail frontier closed models, though the gap has narrowed considerably since 2023.<\/figcaption><\/figure>\n\n\n\n

Open-weight models (often called open-source) currently trail proprietary frontier models by roughly 6 to 12 months on capability benchmarks.[5]<\/a><\/sup> This means any capability a closed model has today, including Mythos-class cyber capability, we should expect to see in an open-weight model within a year. The implications for cybersecurity, CBRN risk, and financial and cyber fraud are serious.<\/p>\n\n\n\n

Open-weight models can be released with safeguards, but those safeguards are trivially easy to strip off through fine-tuning.[6]<\/a><\/sup> There is currently no known technical method for releasing open weights in a way that prevents downstream fine-tuning for cyberattacks, bioweapon assistance, or terrorism support. In practice, the capabilities of the best open-weight models should be treated as uncontrollable once released.<\/p>\n\n\n\n

\u201cPowerful agentic systems are going to be built, because they\u2019ll be more useful, economically more useful, scientifically more useful. But then those systems become even more powerful in the wrong hands, too.\u201d<\/p>\u2014 Demis Hassabis, CEO of Google DeepMind, Paris AI Action Summit, February 2025[10]<\/a><\/sup><\/cite><\/blockquote>\n\n\n\n

4. Current Frontier Models Can Substantially Uplift CBRN Threats<\/h2>\n\n\n\n
\"Anthropic
Anthropic long-form virology task evaluations across two task suites and three sub-components (sequence design, protocol design, end-to-end). Claude Sonnet 4.6, Opus 4.6, and Mythos Preview all score between 0.79 and 1.00, indicating near-saturation on these long-form biology workflows.<\/figcaption><\/figure>\n\n\n\n

Frontier models are now reaching the point where they can meaningfully uplift malicious actors attempting to create novel CBRN threats.[7]<\/a><\/sup> Major model providers invest heavily in safeguards to prevent this, but open-weight models remain only 6 to 12 months behind. Within the next two years, fully open-weight models will likely reach a level where an unskilled actor\u2019s ability to develop a novel biological or chemical threat is materially improved.<\/p>\n\n\n\n

OpenAI\u2019s evaluation of GPT-4.5 under its Preparedness Framework found comparable results, rating the model at the boundary of its \u201cMedium\u201d threshold for biological risk.[12]<\/a><\/sup> Expert evaluators\u2014credentialed virologists and biosecurity researchers\u2014found that the model could substantially accelerate the information-gathering and protocol-design phases of biological threat development, reducing the time required for a moderately skilled actor to move from intent to actionable synthesis plan. OpenAI noted that while its safeguards reduced compliance on direct requests, adversarial elicitation techniques could recover a significant fraction of the underlying capability. The consistent finding across Anthropic and OpenAI evaluations is that the knowledge required to design novel CBRN threats is now latent in frontier model weights\u2014not as a deliberate design choice, but as an emergent consequence of training on the world\u2019s scientific literature. Safeguards reduce willingness; they do not reduce capability.<\/p>\n\n\n\n

\u201cAs of mid-2025, our measurements show that LLMs may already be providing substantial uplift in several relevant areas, perhaps doubling or tripling the likelihood of success. We believe that models are likely now approaching the point where, without safeguards, they could be useful in enabling someone with a STEM degree but not specifically a biology degree to go through the whole process of producing a bioweapon.\u201d<\/p>\u2014 Dario Amodei, CEO of Anthropic, \u201cThe Adolescence of Technology,\u201d January 2026[11]<\/a><\/sup><\/cite><\/blockquote>\n\n\n\n

Concluding Thoughts<\/h2>\n\n\n\n

Taken together, these four trends are an urgent call to action for global policy makers. The time to regulate is before a major cyber or biosecurity incident, not after.<\/p>\n\n\n\n