The short answer
AI safety is the work of making sure AI systems do what people intend without causing harm, from wrong answers and data leaks to misuse for fraud or cyberattacks and risks to society. The International AI Safety Report 2026, published on 3 February 2026 by more than 100 experts chaired by Yoshua Bengio, groups the risks into misuse, malfunctions and systemic risks, and finds that safeguards still fail at a moderately high rate.
| Plain meaning | Keeping AI systems useful, reliable and hard to misuse |
|---|---|
| Three kinds of risk | Misuse, malfunctions and systemic risks (International AI Safety Report 2026) |
| Key report | International AI Safety Report 2026, 3 February 2026, chair Yoshua Bengio |
| Rules in force | EU AI Act: prohibitions since 2 February 2025, most rules applicable from 2 August 2026 |
| At work | Checking outputs, protecting data, limiting what agents can do, training staff |
What AI safety means, in plain English
AI safety covers everything that can go wrong when people build and use AI, and the methods for preventing it. Some of it is about today's tools: a model that invents a figure, an assistant that leaks a confidential file, an agent tricked into sending email. Some of it is about the future: what happens as systems become more capable and more autonomous. The field spans technical research, company policies, testing and law.
You will also hear related terms. AI security usually means protecting AI systems from attackers, as with prompt injection. Responsible AI usually means fairness, transparency and accountability in how organisations use it. In practice they overlap, and a sensible workplace approach covers all three.
The three kinds of AI risk
The International AI Safety Report 2026 is the most useful single map of the field. It was published on 3 February 2026, chaired by Yoshua Bengio, written with more than 100 independent experts and backed by an advisory panel nominated by more than 30 countries and international organisations. It sorts risks into three groups.
- Misuse: people deliberately using AI to cause harm, such as fraud, cyberattacks, manipulation, or help with biological or chemical weapons.
- Malfunctions: AI failing or behaving unexpectedly without anyone intending it, from unreliable answers to scenarios where people lose control of a system.
- Systemic risks: effects of widespread use, such as impacts on the labour market and on human autonomy.
What the 2026 report found
- Tests miss things. The report describes an evaluation gap: tests before release often fail to predict how a model behaves in the real world, and models increasingly tell test settings apart from real use, so dangerous capabilities could go undetected. Our benchmark explainer covers why scores mislead.
- Safeguards can be bypassed. Prompt injection success rates have fallen but remain relatively high, harmful requests can sometimes be split into harmless-looking steps, and watermarks can often be removed.
- Capabilities move fast. Some coding tasks that take a person about half an hour can now be done reliably by AI agents, up from under 10 minutes a year earlier, and the length of tasks AI can complete has doubled roughly every seven months.
- Use is widespread. At least 700 million people use leading AI systems every week.
- Rules are mostly voluntary. In 2025, 12 companies published or updated frontier AI safety frameworks, but most risk management remains voluntary, and open-weight models cannot be recalled once released.
Why AI safety matters at work
Most teams will never train a model, but every team that uses one inherits its risks. The ones that show up at work are concrete.
- Wrong answers presented confidently. Hallucinations end up in reports and client emails when nobody checks. See how to check if AI is telling the truth and how to reduce hallucinations.
- Data leaving the building. Personal plans of ChatGPT, Claude and Gemini can use chats for training unless you change a setting, while business plans generally do not by default. Our guide to keeping your data private in AI tools lists the exact switches.
- Agents acting beyond their brief. As assistants gain computer use and connectors, a mistake or an injected instruction can send, delete or buy. Anthropic's own help centre warns that Claude can make mistakes and occasionally act outside its intended boundaries.
- Legal duties. In the EU, the AI Act's ban on practices such as emotion recognition in workplaces has applied since 2 February 2025, and supervision and enforcement of AI literacy and other rules began on 2 August 2026. See our EU AI Act explainer.
The report's own answer to imperfect safeguards is defence in depth: layer several protections, because no single one reliably prevents harm. For a team, that means approved tools on business plans, a short policy, human review of anything that leaves the company, minimum permissions for agents and a way to report incidents. Start with using AI at work safely and writing an AI policy for your company.
A 15-minute safety check you can try
Pick one way you or your team uses AI every week, such as drafting client emails or summarising contracts, and answer five questions.
- What goes in? List the data you paste or connect. Is any of it personal, confidential or regulated?
- Which plan and setting? Is it a business plan, or a personal one with training switched off?
- What comes out, and who checks it? Name the person who reviews the output before it is used.
- What can it do? Can the tool only answer, or can it send, edit or buy? Remove any permission the task does not need.
- What if it is wrong? Describe the worst realistic mistake and how you would notice.
Paste your answers into the first prompt below for a second opinion. If any answer is "I don't know", that is your first fix.
Related terms worth knowing
- Prompt injection: the top security risk for AI applications.
- Hallucination: confident but wrong output, the most common malfunction at work.
- EU AI Act: the main law setting safety duties in Europe.
- Benchmark: tests that measure capability, with known limits.
- AGI: the more speculative long-term debate within AI safety.
- Open-source model: weights anyone can download, which raises specific safety trade-offs.
Common misconceptions about AI safety
- "AI safety is only about robots taking over." Long-term risks are part of the field, but most of the 2026 report covers present-day problems such as fraud, cyber misuse, unreliable outputs and labour effects.
- "The AI company handles safety for me." Makers build safeguards, but the report says attackers still succeed at a moderately high rate. How you configure and supervise tools is your part.
- "Passing the tests means it is safe." The report's evaluation gap finding says pre-release tests often do not predict real-world behaviour.
- "Safety slows us down." A short policy, the right plan and a review step take hours to set up and save the cost of one leaked file or wrong figure.
- "It is all regulated now." The EU AI Act sets real duties, but the report notes that most risk management by AI developers remains voluntary.
Prompts to copy
Review one AI workflow for risk
Here is how my team uses AI for [task]: data in: [list], tool and plan: [name], who checks the output: [name or nobody], what the tool can do: [answer only / send / edit / buy], worst realistic mistake: [describe]. Using the three risk groups of misuse, malfunction and systemic risk, list the top three risks in this workflow and one practical fix for each that takes less than a day.
You get: A ranked list of risks with quick, practical fixes
Draft a one-page AI use rule
Draft a one-page AI use rule for a [size] team in [industry] in [country]. Cover approved tools and plans, what data must never be entered, when a human must check output, which actions an AI agent may never take without approval, and how to report a problem. Plain English, no legal jargon, under 350 words.
You get: A short rule your team can actually follow