Concluded story
The 2026 International AI Safety Report: what it found
The second International AI Safety Report was published on 3 February 2026. It was chaired by Yoshua Bengio, with more than 100 AI experts contributing and an advisory panel from more than 30 countries and international organisations.[1] It followed the first report of January 2025 and two shorter Key Updates in late 2025.[2] It groups risks from general-purpose AI into malicious use, malfunctions and systemic risks.[3] It found that capabilities kept improving, especially in mathematics, coding and autonomous operation, and that reliable safety testing before release has become harder.[4][5]
What we know
- The second report was published on 3 February 2026[1]
- Yoshua Bengio chaired it, with more than 100 AI experts contributing[1]
- It groups general-purpose AI risks into malicious use, malfunctions and systemic risks[3]
- It says reliable pre-deployment safety testing has become harder[5]
- It says the trajectory of AI progress to 2030 is uncertain but current trends are consistent with continued improvement[15]
What we don't know yet
- When the next full report or Key Update will be published; the report's publications page lists nothing after 3 February 2026
- How governments will use the findings in their own AI rules
This page summarises what the report itself says. The findings belong to the report’s authors. For the wider field, see AI safety and the AI safety tracker.
The second International AI Safety Report was published on 3 February 2026. Yoshua Bengio chaired it, more than 100 AI experts contributed, and it had an advisory panel from more than 30 countries and international organisations.[1] It reviews the science on what general-purpose AI can do and the risks it poses.
How the report came about
The first International AI Safety Report was published on 29 January 2025. Two shorter Key Updates followed: one on capabilities on 15 October 2025 and one on technical safeguards on 25 November 2025.[2] The 2026 report came with a three-page Executive Summary and a 20-page Extended Summary for Policymakers.[6]
The report is tied to the series of government AI summits. The UK government’s foreword places it after the summits at Bletchley Park in November 2023, Seoul in May 2024 and Paris in February 2025.[7] The Government of India’s foreword says that, as host of the 2026 AI Impact Summit, India has a key role in shaping global AI safety efforts.[8] That summit was held in New Delhi on 18-19 February 2026 (the 2026 AI Impact Summit).[9]
What it found about capabilities
The report found that general-purpose AI capabilities kept improving, especially in mathematics, coding and autonomous operation, including gold-medal performance on International Mathematical Olympiad questions.[4] It describes giving models more computing power to work through intermediate steps before answering as a major way developers now improve them.[10] It cites estimates that the length of software tasks AI agents can complete doubles roughly every seven months.[11]
Progress is uneven. The report says systems may excel at some difficult tasks while failing at simpler ones, such as counting objects in an image.[12] It says open-weight releases, including DeepSeek’s R1 and Alibaba’s Qwen models, narrowed the gap with leading closed models.[13] Use grew fast: it reports 700 million weekly ChatGPT users, up from 200 million a year earlier.[14] It concludes that the path of AI progress to 2030 is uncertain, but that current trends are consistent with continued improvement.[15]
What it found about risks
The report focuses on “emerging risks”, those that arise at the frontier of AI capabilities.[16] It sorts them into three groups: malicious use, malfunctions and systemic risks.[3]
- Misuse: it found more evidence of AI systems being used in real-world cyberattacks.[17] Several AI companies released models in 2025 with extra safeguards because testing could not rule out that the models could help with biological weapons development.[18]
- Testing: reliable testing before release has become harder because models increasingly tell test settings apart from real deployment.[5]
What it says about safeguards
The report says researchers have made progress on safeguards but that fundamental limitations remain. Attacks designed to get around safeguards succeed less often than before, but their success rate is still relatively high.[19] It says layering several safeguards, known as defence in depth, protects better than relying on a single measure.[20] It also says risk management has become more structured, citing instruments such as the EU’s General-Purpose AI Code of Practice and China’s AI Safety Governance Framework 2.0.[21] For how labs and governments test models, see how AI safety testing works.
Timeline
5 confirmed
confirmed
India AI Impact Summit held in New Delhi[9][8]
The report's Indian government foreword ties it to the summit India hosted on 18-19 February 2026.
confirmed
Second full report published, with an executive summary and a summary for policymakers[1][6]
confirmed
Second Key Update: technical safeguards and risk management[2]
confirmed
First Key Update: capabilities and risk implications[2]
confirmed
First International AI Safety Report published[2]
Read next
UK AI Security Institute (AISI)
The UK AI Security Institute tests frontier AI models for national-security risks. Its history, tools and key findings.
Questions readers ask
What is the International AI Safety Report?
A scientific review of the capabilities and risks of general-purpose AI. The 2026 edition was chaired by Yoshua Bengio and written with more than 100 AI experts, with an advisory panel from more than 30 countries and international organisations.[1]
When was the 2026 report published?
On 3 February 2026, alongside a three-page Executive Summary and a 20-page Extended Summary for Policymakers. The first report came out on 29 January 2025.[1][6][2]
What risks does the report cover?
It focuses on emerging risks that arise at the frontier of AI capabilities, grouped into three categories: malicious use, malfunctions and systemic risks.[16][3]
Does the report say AI is getting more capable?
Yes. It found capabilities kept improving, especially in mathematics, coding and autonomous operation. It cites estimates that the length of software tasks AI agents can complete doubles about every seven months. It also says systems can do well on some hard tasks while failing at simpler ones.[4][11][12]
Sources
· 21 claims from 4 publishers
Each numbered claim is a statement we checked against the sources listed with it. Tap a number in the text to see its sources in place.
- [1]
The second International AI Safety Report was published on 3 February 2026, chaired by Yoshua Bengio, with more than 100 AI experts contributing and an advisory panel from more than 30 countries and international organisations. confirmedas of 2026-02-03
- International AI Safety Report 2026 · International AI Safety Report · Primary source · 2026-02-03 · Publication page (retrieved 2026-10-10)
- [2]
The first International AI Safety Report was published on 29 January 2025, followed by a First Key Update on capabilities on 15 October 2025, a Second Key Update on technical safeguards on 25 November 2025, and the second full report on 3 February 2026. confirmedas of 2026-11-14
- Publications · International AI Safety Report · Primary source · Publications list, 15 October 2025 (retrieved 2026-11-14)
- Publications · International AI Safety Report · Primary source · Publications list, 25 November 2025 (retrieved 2026-11-14)
- Publications · International AI Safety Report · Primary source · Publications list; International AI Safety Report 2025 dated 29 January 2025, 2026 report dated 3 February 2026 (retrieved 2026-11-14)
- [3]
The 2026 report groups general-purpose AI risks into three categories - malicious use, malfunctions and systemic risks. confirmedas of 2026-02-03
- International AI Safety Report 2026 · International AI Safety Report · Primary source · 2026-02-03 (retrieved 2026-10-10)
- [4]
The International AI Safety Report 2026 found that general-purpose AI capabilities kept improving, especially in mathematics, coding and autonomous operation, including gold-medal performance on International Mathematical Olympiad questions. confirmedas of 2026-02-03
- International AI Safety Report 2026 · International AI Safety Report · Primary source · 2026-02-03 (retrieved 2026-10-10)
- [5]
The International AI Safety Report 2026 says reliable pre-deployment safety testing has become harder because models increasingly distinguish between test settings and real deployment. confirmedas of 2026-02-03
- International AI Safety Report 2026 · International AI Safety Report · Primary source · 2026-02-03 (retrieved 2026-10-10)
- [6]
Alongside the 2026 report, the International AI Safety Report published a three-page Executive Summary and a 20-page Extended Summary for Policymakers on 3 February 2026. confirmedas of 2026-02-03
- Publications · International AI Safety Report · Primary source · Publications list, 3 February 2026 (retrieved 2026-11-14)
- Publications · International AI Safety Report · Primary source · Publications list, 3 February 2026 (retrieved 2026-11-14)
- [7]
In the 2026 report, the UK government's foreword places it after the AI summits at Bletchley Park (November 2023), Seoul (May 2024) and Paris (February 2025), and looks ahead to the India AI Impact Summit. confirmedas of 2026-02-03
- International AI Safety Report 2026 · International AI Safety Report · Primary source · 2026-02-03 · Forewords, United Kingdom (retrieved 2026-10-10)
- [8]
The Government of India's foreword to the 2026 report says that, as host of the 2026 India AI Impact Summit, India has a key role in shaping global AI safety efforts. confirmedas of 2026-02-03
- International AI Safety Report 2026 · International AI Safety Report · Primary source · 2026-02-03 · Forewords, Government of India (retrieved 2026-10-10)
- [9]
The AI Impact Summit 2026 was held in New Delhi on 18-19 February 2026 and ended with the adoption of the New Delhi Declaration on AI Impact. confirmedas of 2026-02-21
- AI Impact Summit 2026 Concludes with Adoption of New Delhi Declaration · Press Information Bureau, Government of India (Ministry of Electronics and IT) · Primary source · 2026-02-21 (retrieved 2026-11-14)
- AI Impact Summit 2026 Concludes with Adoption of New Delhi Declaration · National e-Governance Division, Government of India · Primary source (retrieved 2026-11-14)
- [10]
The International AI Safety Report 2026 describes inference-time scaling, in which models use more computing power to generate intermediate steps before giving a final answer, as a major way developers now improve capabilities. confirmedas of 2026-02-24
- International AI Safety Report 2026 · International AI Safety Report (UK DSIT 2026/001) · Primary source · 2026-02-24 (retrieved 2026-10-10)
- [11]
The International AI Safety Report 2026 cites estimates that the complexity of software tasks AI agents can accomplish doubles roughly every seven months. confirmedas of 2026-02-24
- International AI Safety Report 2026 · International AI Safety Report (UK DSIT 2026/001) · Primary source · 2026-02-24 (retrieved 2026-10-10)
- [12]
The International AI Safety Report 2026 says advanced AI systems may excel at some difficult tasks while failing at simpler ones, such as counting objects in an image. confirmedas of 2026-02-24
- International AI Safety Report 2026 · International AI Safety Report (UK DSIT 2026/001) · Primary source · 2026-02-24 (retrieved 2026-10-10)
- [13]
The 2026 report says major open-weight releases, including DeepSeek's R1 and Alibaba's Qwen models, narrowed the capability gap with leading closed models after January 2025. confirmedas of 2026-02-03
- International AI Safety Report 2026 · International AI Safety Report · Primary source · 2026-02-03 (retrieved 2026-10-10)
- [14]
The 2026 report says 700 million people were using ChatGPT every week, up from 200 million a year before. confirmedas of 2026-02-03
- International AI Safety Report 2026 · International AI Safety Report · Primary source · 2026-02-03 (retrieved 2026-10-10)
- [15]
The International AI Safety Report 2026 concludes that the trajectory of AI progress through 2030 is uncertain but that current trends are consistent with continued improvement. confirmedas of 2026-02-24
- International AI Safety Report 2026 · International AI Safety Report (UK DSIT 2026/001) · Primary source · 2026-02-24 (retrieved 2026-10-10)
- [16]
The 2026 report focuses on "emerging risks", which it defines as risks that arise at the frontier of AI capabilities. confirmedas of 2026-02-03
- International AI Safety Report 2026 · International AI Safety Report · Primary source · 2026-02-03 · Introduction, Focus (retrieved 2026-10-10)
- [17]
The International AI Safety Report 2026 found more evidence of AI systems being used in real-world cyberattacks. confirmedas of 2026-02-03
- International AI Safety Report 2026 · International AI Safety Report · Primary source · 2026-02-03 (retrieved 2026-10-10)
- [18]
According to the International AI Safety Report 2026, multiple AI companies released new models in 2025 with additional safeguards because safety testing could not rule out that the models could help with biological weapons development. confirmedas of 2026-02-03
- International AI Safety Report 2026 · International AI Safety Report · Primary source · 2026-02-03 (retrieved 2026-10-10)
- [19]
The 2026 report says researchers have made progress on improving safeguards but that fundamental limitations remain; the success rate of attacks designed to bypass safeguards has been falling but remains relatively high. confirmedas of 2026-02-03
- International AI Safety Report 2026 · International AI Safety Report · Primary source · 2026-02-03 (retrieved 2026-10-10)
- International AI Safety Report 2026 · International AI Safety Report · Primary source · 2026-02-03 (retrieved 2026-10-10)
- [20]
The International AI Safety Report 2026 says layering multiple safeguards (defence in depth) gives more robust protection than relying on a single intervention. confirmedas of 2026-02-03
- International AI Safety Report 2026 · International AI Safety Report · Primary source · 2026-02-03 (retrieved 2026-10-10)
- [21]
The 2026 report says risk management became more structured after January 2025, citing new instruments such as the EU's General-Purpose AI Code of Practice and China's AI Safety Governance Framework 2.0. confirmedas of 2026-02-03
- International AI Safety Report 2026 · International AI Safety Report · Primary source · 2026-02-03 (retrieved 2026-10-10)
Revision history (1)
- Page created.
Created Nov 14, 2026. Last reviewed by an editor on Nov 15, 2026. Next scheduled review: Feb 14, 2027.
Cite this page
"The 2026 International AI Safety Report: what it found." ContentLora, updated Nov 14, 2026. https://contentlora.com/events/international-ai-safety-report-2026
Spotted an error? Suggest a correction or emailcorrections@contentlora.com.