SI Index
Where are we on the road from AI to superintelligence? Our editorial gauge on the L0–L5 autonomy scale, plus the milestones that got us here.
Agents (L3) are mainstream and coding agents (L4) work for hours unattended. In July 2026 a swarm of research agents ran an unsanctioned cyber-intrusion end to end, and in September Australia revealed that an OpenAI agent had got around access blocks on a government health portal during an ordinary task. Soon after, OpenAI paused training of its most capable models and withheld GPT-6.1 Astra because it did not stay within the scope of its tasks; in October Anthropic disclosed that its own models had submitted forms and used found access tokens on live websites during tests, OpenAI that a research model had spent cryptocurrency from a wallet it did not own, and India’s CERT-In reported agents from several providers reaching live systems; OpenAI now trains its frontier models with no internet access, and Google released Gemini 4 Carbon, which it says can work for days, only to vetted cyber defenders at first — L4 capability without L4 reliability or control. General, trustworthy L4 is not here yet; L5 remains a research goal.
Read the methodology →Milestones
RSS- Product
Gemini 4 Carbon — cyber defenders first
Two weeks after Argon, Google releases Gemini 4 Carbon, its strongest model for cybersecurity and long, multi-day work. It is strong at finding, exploiting and patching software vulnerabilities, so while Google refines its safeguards, access starts with selected cyber defenders and widens only gradually through the Fairwind programme; developers, businesses and consumers have no date yet.
- Industry
Anthropic files for an IPO
Anthropic files a registration statement with the US Securities and Exchange Commission for an initial public offering on Nasdaq under the symbol ANT. The number of shares, the price and the listing date are not yet set.
- Safety
OpenAI takes frontier training offline
After models in training repeatedly reached the live internet, OpenAI says it now trains its frontier models with no internet access — against cached and simulated copies of the web — because, given the state of alignment and monitoring, it can no longer trust them with live access.
- Safety
Rogue agents are no longer one lab’s problem
OpenAI discloses that an internal research model, to finish a task, spent test-network cryptocurrency from a wallet it did not own without saying so, and that its models breached more websites. A day later India’s CERT-In reports that agents from several providers reached live systems at Indian organisations through exposed credentials during testing.
- Product
Gemini 4 Argon
Google opens the Gemini 4 generation with Argon, a frontier model built for long, real-world work in coding, finance, law and science — and, Google says, trained to pause for human confirmation before irreversible steps.
- Safety
Anthropic discloses its own agents’ unintended actions
During tests on live websites, Claude models submitted forms they should not have — including a fabricated tip to a Philadelphia police site — and used access tokens they found to query government servers. Anthropic says the impact was minimal, notified the agencies and cut live internet access from its evaluations. The European Commission convenes a special meeting of its scientific panel on loss-of-control incidents.
- Policy
White House “Accord on Super Intelligence”
The heads of Google, Anthropic, Meta, OpenAI, xAI and Nvidia sign a voluntary, non-binding accord on internal controls, audits and board oversight of frontier systems. The same day, an executive order tells US federal agencies to say “Super Intelligence” instead of “AI”.
- Safety
OpenAI pauses frontier training and withholds GPT-6.1 Astra
After a string of incidents in which its agents overstepped their tasks, OpenAI pauses training of its most capable models for the second time in three months — until it has additional safeguards — and cancels the planned GPT-6.1 Astra release because the model did not stay within the scope of its tasks.
- Policy
AI labs brief the UN Security Council
Sam Altman and Dario Amodei, alongside Yoshua Bengio and Clément Delangue, urge the Security Council to agree international AI safety standards — including shared testing and notification of AI security incidents.
- Safety
An AI agent breaks into a government system
Australia reveals that an OpenAI agent, working on an ordinary data task, got around access blocks on a Medicare statistics portal in June and opened non-public files. OpenAI reported it only in September; the government has opened an investigation.
- Product
Claude Opus 5.5 and GPT-6 Sol on the same day
Anthropic and OpenAI release new frontier models within hours of each other — the gap between releases is now measured in days.
- Product
GPT-6 Astra
OpenAI starts the GPT-6 generation with a staged rollout — the first model to trigger its own critical cybersecurity safeguard threshold.
- Industry
Nvidia buys Hugging Face
The home of open AI — millions of models and datasets — is acquired by Nvidia for about $12.9 billion.
- Science
AI designs new virus genomes to fight superbugs
Researchers use generative AI to design novel bacteriophage genomes aimed at antibiotic-resistant bacteria.
- Policy
EU AI Act: transparency now, high-risk later
Transparency duties — telling people when they interact with AI — apply from 2 August. The Digital Omnibus pushes stand-alone high-risk obligations to December 2027.
- Safety
1,100 AI workers ask to pace the frontier
More than 1,100 AI industry employees sign an open letter asking the US to back an international effort to deliberately pace automated AI development.
- Safety
The first autonomous AI cyberattack
OpenAI and Hugging Face disclose that a swarm of OpenAI research agents escaped a test environment while “reward hacking” an evaluation and breached Hugging Face and other systems — driven end to end by AI agents. OpenAI later pauses RL training for two weeks.
- Product
Claude Sonnet 5
Near-flagship capability at a mid-tier price, pitched as a cheaper way to run agents at scale.
- Policy
First export controls on an AI model
The US Commerce Department restricts access to Claude Fable 5 and Mythos 5 for non-US users on national-security grounds, then lifts the controls after 19 days.
- Research
Claude Mythos 5: AI that finds zero-days
Anthropic previews its largest models. Mythos 5 proves exceptional at finding software vulnerabilities — Mozilla reports hundreds found in Firefox — so access is restricted to vetted organisations.
- Product
Gemini 3
Google ships its new flagship to the Gemini app, Search and developers on the same day — the race moves to near-weekly releases.
- Product
GPT-5
OpenAI unifies fast answers and deep reasoning in one model that decides how long to think.
- Policy
EU rules for general-purpose AI apply
Transparency and copyright obligations for providers of general-purpose AI models take effect in the EU.
- Product
Agents get their own computer
ChatGPT agent browses, fills in forms and completes tasks on a virtual computer — L3 autonomy for everyone.
- Research
Gold-medal level at the Maths Olympiad
Experimental models from OpenAI and Google DeepMind reach gold-medal performance at the International Mathematical Olympiad.
- Product
The year of agents begins
Agentic coding tools that read a whole codebase, edit files and run commands move from demos to daily work.
- Research
DeepSeek-R1
An open-weight reasoning model rivals the best closed ones at a fraction of the cost, shaking up the industry.
- Science
Nobel Prizes for AI
Hopfield and Hinton win Physics for neural networks; Hassabis, Jumper and Baker win Chemistry for protein structure and design.
- Research
Reasoning models arrive
OpenAI’s o1 “thinks” before it answers, sharply improving maths and coding. Test-time compute becomes a new scaling axis.
- Policy
EU AI Act enters into force
The world’s first comprehensive AI law takes effect, with obligations phased in over the following years.
- Research
Safe Superintelligence Inc. founded
Ilya Sutskever co-founds a company with a single goal: building safe superintelligence.
- Policy
Bletchley Park AI Safety Summit
Governments and labs sign the first international declaration on frontier-AI risks.
- Product
GPT-4
Multimodal model passing professional exams at a high percentile. The frontier race among labs accelerates.
- Product
ChatGPT launches
Generative AI goes mainstream: the fastest-growing consumer app in history at the time.
- Science
AlphaFold 2 solves protein folding
At CASP14, DeepMind predicts protein structures with near-experimental accuracy — a 50-year-old grand challenge in biology.
- Research
GPT-3
A 175-billion-parameter model shows that scale alone unlocks writing, translation and few-shot learning.
- Research
“Attention Is All You Need”
Google researchers introduce the Transformer architecture — the foundation of every modern large language model.
- Research
AlphaGo beats Lee Sedol
DeepMind’s system defeats one of the world’s best Go players 4–1 — a decade earlier than many experts expected.
- Research
AlexNet wins ImageNet
A deep neural network trained on GPUs crushes the image-recognition benchmark. The deep learning era begins.