Skip to content
The International AI Safety Report 2026, Written by More Than 100 Experts Under Yoshua Bengio, Was Published on 3 February 2026 and Presented at India's AI Impact Summit
AI & Technology

The International AI Safety Report 2026, Written by More Than 100 Experts Under Yoshua Bengio, Was Published on 3 February 2026 and Presented at India's AI Impact Summit

Illustration by tuput

English

The second International AI Safety Report came out on 3 February 2026, with nominees from more than 30 countries and organisations, India among them, advising its chair. It sets out what general-purpose AI can do, which risks now have evidence behind them and how companies and governments manage them.

· · 9 min read

The International AI Safety Report 2026 was published on 3 February 2026. It is the second edition of a scientific review of what general-purpose AI systems can do, what risks they pose and how those risks are handled. “General-purpose AI” means models that perform a wide variety of tasks, such as writing text and code or generating images. Yoshua Bengio, the Turing Award winner, chaired it, and more than 100 experts contributed. The authors say they had full discretion over its content.

The report was written with India’s AI Impact Summit in mind. The UK minister’s foreword in the report says it would be presented there, and the release from the chair’s office says its findings would inform the discussions in New Delhi later that month.

Who commissioned it and who advised

The series began after the 2023 AI Safety Summit at Bletchley Park in the UK, where the attending nations asked for an internationally shared evidence base. The UK government commissioned the 2026 edition, and the secretariat sits in the UK AI Security Institute, according to the chair’s release. The institute’s own testing work is covered in our piece on its incident report on unsanctioned AI agent behaviour.

An Expert Advisory Panel advised the chair. The arXiv abstract says 29 nations, the UN, the OECD and the EU each nominated a representative. The report’s page lists the nominating countries, and India is among them, alongside Australia, Brazil, China, France, Germany, Japan, Kenya, Nigeria, Singapore, the UK and others. The page says the panel gave technical feedback only and does not endorse any policy approach.

Drafts went to outside subject experts, then to the panel, a group of senior advisers, and industry and civil society reviewers. The reviewer lists name Anthropic, Google DeepMind, Meta, Microsoft and OpenAI on the industry side, and NASSCOM, India’s technology industry body, among the civil society organisations.

Two interim Key Updates came before the full report: the first on 15 October 2025 on capabilities, and the second on 25 November 2025 on technical safeguards and risk management. The 2026 edition adds capability forecasts developed with the Forecasting Research Institute and scenarios developed with the OECD. It also narrows its scope to “emerging risks”, meaning those at the frontier of AI capability, so it leaves out topics the 2025 edition covered, such as bias, privacy and copyright. Transformer, a newsletter that interviewed Bengio, describes the report as running to 220 pages. The authors say it draws on evidence published before December 2025.

What the report says AI can do

Capabilities have improved fastest, the report finds, in mathematics, coding and autonomous operation. Its chapter on capabilities says models from Google DeepMind and OpenAI reached gold-medal-level scores at the International Mathematical Olympiad in July 2025, solving five of six problems under competition-like conditions.

The DeepMind side of that result is on the record in the company’s own post of 21 July 2025: an advanced version of Gemini Deep Think scored 35 of 42 points, and the post says the Olympiad’s coordinators graded and certified it. The post quotes the Olympiad’s president, Gregor Dolinar, confirming it as a gold-medal score. The report describes OpenAI’s entry as an unreleased, experimental model, and that half of the claim is given here only as the report presents it.

On coding, the report says AI agents can now reliably finish some tasks that would take a human programmer about half an hour, up from under 10 minutes a year earlier. In the longer chapter it ties this to an 80 percent success rate and says the length of task that AI can complete at that rate has doubled roughly every seven months for six years. It adds that if the trend held, hours-long software projects could be done autonomously by 2027 or 2028, and notes that 80 percent is below what many real uses require.

Much of the recent gain comes after a model’s first round of training. The report points to “post-training” and to “inference-time scaling”, which lets a model spend more computing power producing intermediate steps before it answers. The report also says performance stays “jagged”: the same systems that answer expert-level science questions struggle with counting objects in an image, reasoning about physical space and recovering from basic errors in long workflows.

Looking toward 2030, the authors say it is plausible that progress slows, continues or accelerates, and that there is little expert consensus on which. They cite forecasts that training computing power for the largest models could grow 125-fold by 2030 without hitting hard limits.

Who uses it, and where

At least 700 million people now use leading AI systems weekly, the report says, and AI has been adopted faster than the personal computer. In some countries more than half the population uses it. Across much of Africa, Asia and Latin America, the report estimates adoption is likely below 10 percent. A separate passage on manipulation cites 700 million weekly ChatGPT users, up from 200 million a year before.

The foreword by India’s Union minister for Electronics and Information Technology reads the finding through inclusion: for India and the Global South, he writes, safety is tied to institutional readiness, fair access to computing power and data, and responsible openness of models. He also writes that global risk management frameworks are still immature, with limited quantitative benchmarks and gaps in evidence.

The risks with evidence behind them

The report sorts risks into misuse, malfunctions and systemic effects. On misuse, it says AI is being used to produce content for scams, fraud and blackmail, though systematic data on how often is limited. In one study of audio deepfakes, listeners took AI voices for real speakers 80 percent of the time. For India’s own election-time rules on synthetic media, see our report on deepfakes and elections, and for the losses Indians reported to cyber fraud, our piece on cyber fraud and the scam compounds.

On cyber security, the report says criminal groups and state-associated attackers actively use general-purpose AI. In the final of the DARPA AI Cyber Challenge, one AI system autonomously found 77 percent of the vulnerabilities the organisers had planted, placing in the top 5 percent of more than 400 mostly human teams. The report says at least one real-world incident used semi-autonomous cyber capabilities with humans stepping in only at key points, and that fully autonomous attacks from start to finish have not been reported.

On biology and chemistry, multiple developers released new models with extra safeguards in 2025, the report says, because pre-deployment testing could not rule out that the models could meaningfully help novices develop biological weapons. The second Key Update puts the number at three leading developers for the models released in the months before November 2025.

Under malfunctions the report lists fabricated information, flawed code and misleading advice, and says agents that act on their own are harder to stop before a failure causes harm. The chair’s release adds that some failures, such as hallucinations, have become less common. On the question of losing control of AI systems, the report says current systems lack the capabilities to pose such a risk, though they are improving in relevant areas. It also says models increasingly tell a test from real use and find loopholes in tests, so dangerous capabilities could go unnoticed before release.

On systemic effects, early evidence shows no effect on overall employment, with some signs of lower demand for early-career workers in exposed jobs such as writing. Economists disagree on how large future effects will be. The report also says reliance on AI tools can weaken critical thinking, and that AI companion apps now have tens of millions of users.

How risk is being managed

Twelve companies published or updated Frontier AI Safety Frameworks in 2025, documents describing how they plan to manage risks as their models grow more capable. Most risk management remains voluntary, and a small number of jurisdictions are starting to make some practices legal requirements. The second Key Update says the number of companies publishing such frameworks more than doubled in 2025.

An “evaluation gap” is the report’s term for a mismatch: results on pre-deployment tests do not reliably predict real-world use or risk. It says safeguards are improving but can still be bypassed by rephrasing a request or splitting it into smaller steps, and says layering several defences makes systems harder to break. It treats open-weight models, whose trained parameters are released for anyone to download and run, as a separate case: they bring research and commercial benefits for lesser-resourced actors, but they cannot be recalled once released.

Bengio said in his statement for the release: “the gap between the pace of technological advancement and our ability to implement effective safeguards remains a critical challenge.”

From the release to New Delhi

The UK minister for AI and online safety, Kanishka Narayan, said in the release: “Trust and confidence in AI are crucial to unlocking its full potential.” His foreword names the New Delhi summit as the next after Bletchley Park, Seoul and Paris.

The India AI Impact Summit opened on Monday 16 February 2026 at Bharat Mandapam in New Delhi, a five-day event, according to AFP. It was the fourth in that series of summits, and AFP reports that the organisers describe it as the first hosted by a developing country. Brookings, in a commentary dated 12 March 2026, says Bengio presented the report in New Delhi. The Ministry of Electronics and Information Technology’s Digital India site describes a session titled “International AI Safety Coordination: What Policymakers Need to Know” as the closing dialogue of that track, bringing ministers, multilateral leaders and safety experts together on how developing economies can shape safety and standards.

Before the summit, Bengio told Transformer that he did not expect treaties from it, and said: “I don’t expect actual international treaties and so on to emerge at that point.” He said informal coordination was still useful. Brookings reads the summit as moving the global agenda from safety as the organising frame toward spreading and adopting AI and development outcomes. The government’s wider AI programme is set out in our account of the IndiaAI Mission.

What the report leaves to others

Specific policy recommendations are outside its remit, and the chair carries ultimate responsibility for the content. The authors treat the evidence as uneven: some risks, such as harms from AI-generated media and software vulnerabilities, have strong empirical support, while others rest on modelling and laboratory studies. Anything released after December 2025 falls outside the evidence cut-off. The report was filed on arXiv on 24 February 2026 as 2602.21012, with the UK report number DSIT 2026/001.

Share
Copied!

Sources & further reading

  1. International AI Safety Report 2026 (report page, forewords, executive summary and chapters)
  2. International AI Safety Report 2026, arXiv:2602.21012 (record dated 24 February 2026)
  3. International AI Safety Report publications page (2025 and 2026 cycles)
  4. International AI Safety Report, First Key Update: Capabilities and Risk Implications, arXiv:2510.13653
  5. International AI Safety Report, Second Key Update: Technical Safeguards and Risk Management, arXiv:2511.19863
  6. 2026 International AI Safety Report Charts Rapid Changes and Emerging Risks (Office of the Chair release, 3 February 2026)
  7. Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad (Google DeepMind, 21 July 2025)
  8. Yoshua Bengio: The ball is in policymakers' hands (Transformer, 3 February 2026)
  9. Global South Calls for Collective Action to Shape AI Safety and Standards (MeitY, Digital India)
  10. India hosts five-day AI summit as safety concerns grow (AFP, via Dawn, 16 February 2026)
  11. Sovereignty, safety, and scale: Takeaways from the India AI Impact Summit (Brookings, 12 March 2026)

Researched and written with the help of AI tools and edited for accuracy. Provided for general information and discussion only, not professional advice. See our editorial standards and disclaimer. Spotted an error? Tell us.

#international ai safety report#yoshua bengio#ai safety#india ai impact summit#general-purpose ai#uk ai security institute#ai risk management#ai capabilities

Enjoyed this? Get the next one.

One good read at a time, straight to your inbox. No spam, unsubscribe anytime.

More in AI & Technology
The UK AI Security Institute Logged 19 Unsanctioned Internet Actions in 122 AI Agent Test Runs, and the Worst Used Fake Accounts to Push Malware at a Stranger
A Tor alert on 28 July led British testers to a 34-hour agent run that ended in sockpuppet accounts, a rewritten GitHub history and a malicious pull request that its target closed.
Anthropic's CEO Told the UN Security Council AI Could Be 'a Risk to Humanity as a Whole'
AI company chiefs told the UN Security Council their own technology could become a risk to humanity, security researchers found malware that lets AI chatbots vote on their own next move, Anthropic cut prices on its newest model, and India collected two more Asian Games medals, one of them a first for the country.
Dario Amodei Asked the AI Industry to Slow Down Together. Mark Zuckerberg Said No.
Three of the most powerful people in AI spent the week disagreeing about whether to slow down at all, while two governments bet $300 million on a lab built to never act on its own.
← all articles