top of page

AI security vs AI safety vs AI governance: what's the difference?

Oct 29
7 min read
Infographic comparing AI safety, AI security and AI governance, with the question each asks and a real example of each


If you've ever been in a meeting where one person said "AI safety", another said "AI security" and a third said "we need AI governance", and you quietly wondered whether they were all talking about the same thing, you're in good company.

They're not the same thing. They overlap, they depend on each other, and they get mixed up all the time, even by people who write about AI for a living. So here's what they mean.


A short description:


  • AI safety is about stopping an AI system from causing harm by accident. 

  • AI security is about protecting an AI system from people who want to trick, misuse or break it on purpose.

  • AI governance is about who in your organisation decides how AI is used, who is accountable when it goes wrong, and how you prove you're in control.

The easiest way to tell them apart is the question each one asks. Safety: could this hurt someone even if everyone means well? Security: what could someone with bad intentions do with this? Governance: who's in charge here, and can we show it?


What is AI safety?

AI safety is about making sure an AI system behaves the way it should and doesn't cause harm, even when no one intends for it to do anything wrong. It’s about managing risk from inside the system: biased training data, poor testing, a model used for something it wasn't built for, or outputs nobody checks.


Example: the benefits tool that flagged the wrong people. The Department for Work and Pensions uses a machine-learning tool to help decide which Universal Credit advance claims to check for fraud. In December 2024, the Guardian alleged, using documents released under Freedom of Information, that the DWP's own fairness analysis had found the tool picked people for investigation unevenly, based on their age, disability, marital status and nationality.

Nobody was attacking the system, and nobody set out to build something unfair. The DWP points out that a caseworker always makes the final decision. But the tool was still choosing who gets looked at, and it wasn't choosing evenly. That's a safety problem.


What safety looks like in practice: testing for bias before any tools are launched, checking accuracy against real cases, keeping a human in the loop for decisions that affect people, and monitoring the system after it goes live.


What is AI security?


AI security is about protecting an AI system from people who are deliberately trying to misuse it. The risk comes from outside: someone feeding it instructions it shouldn't follow, poisoning the data it learns from, or pulling out information it should keep private.

Example: the video game face that passed an age check. When the Online Safety Act's age checks came into force on 25 July 2025, platforms like Discord started asking UK users for a face scan so AI could estimate their age. On day one, people found they could pass Discord's check by pointing their phone at a character in the video game Death Stranding. PC Gamer tried it and it worked.

Nobody hacked into anything. People deliberately showed the AI something it wasn't built to expect, to get a result they weren't entitled to. That's security. The EU AI Act even names this kind of attack: inputs designed to make an AI system get it wrong.

What security looks like in practice: limiting what an AI tool is allowed to do, testing it the way an attacker would, protecting training data, controlling who has access, and watching for unusual use.


What is AI governance?


AI governance is the way your organisation runs AI: who approves a new tool, who owns it once it's live, what rules people follow when they use it, what gets written down, and who answers for it when something goes wrong. Most of it has nothing to do with code.


Example: fake cases in the High Court. In June 2025, the High Court in London dealt with two cases where lawyers had put made-up case law in front of the court, with AI tools either known or suspected to be the source. The lawyers were referred to their regulators.

The court didn't stop at the individuals. Dame Victoria Sharp, President of the King's Bench Division, called on regulators and the profession's leaders to make sure lawyers understand their duties when they use AI.


The AI inventing cases was a safety issue. Nobody checking, training or supervising how it was used was a governance issue. If your team uses AI and nobody has set the rules, that gap belongs to the business, not just the person who pressed the button.


What governance looks like in practice: a list of the AI tools you use, a named owner for each, an AI use policy your team actually knows about, risk assessments, and records you can show when someone asks.


AI safety vs security vs governance at a glance


AI safety

AI security

AI governance

The question it asks

Could this cause harm even if everyone means well?

What could someone with bad intentions do with this?

Who's in charge, and can we prove it?

Where the risk comes from

Inside the system: data, design, testing

Outside the system: attackers and misuse

The organisation: unclear ownership, missing rules

Typical problems

Bias, wrong answers, unreliable results

Prompt injection, data poisoning, data leaks

No owner, no policy, no records

Who usually leads it

Product, data and AI teams

IT and security teams

Leadership, legal, risk and compliance

Real example

The DWP's benefits fraud tool

Age checks fooled by a video game

Fake AI case law in the High Court


How they fit together


Think of a new AI tool you're about to roll out, say a chatbot that answers customer questions.

  • Safety makes sure it gives correct, fair answers.

  • Security makes sure nobody can trick it into giving away discounts or customer data.

  • Governance makes sure someone signed it off, someone owns it, there are rules for it, and there's a paper trail if anything goes wrong.

You need all three. A tool can be well protected from attackers and still treat people unfairly, and a well-tested tool can still be tricked. Governance is what makes sure someone is actually checking for both, which is why it usually sits on top.


Where the EU AI Act and ISO 42001 come in


The EU AI Act touches all three. Article 15 says high-risk AI systems must reach an appropriate level of accuracy, robustness and cybersecurity throughout their lifecycle. Accuracy and robustness are safety. Cybersecurity is security, and the Act names specific attacks like data poisoning and adversarial examples. Around that sit requirements on risk management, human oversight and record-keeping, which is governance.

On timing: the Digital Omnibus on AI, in force since 27 July 2026, moved the deadline for most high-risk systems (things like hiring, credit scoring and education) to 2 December 2027, and to 2 August 2028 for AI built into regulated products. It's a delay, not a cancellation. Other parts of the Act, including the Article 50 transparency rules, already apply.

If you're a UK business the Act can apply to you if you sell an AI system into the EU, or if your AI's output is used there, even with no EU office.

ISO/IEC 42001 is where governance gets practical. Published in December 2023, it's the international standard for AI management systems: how you set policies, assign responsibility, assess risk and keep improving. It's voluntary, it doesn't replace the law, and certification is done by independent certification bodies. What it does give you is a structure that makes it much easier to show you're meeting your legal obligations.

Frequently asked questions

Is AI safety the same as AI security? No. AI safety is about preventing accidental harm, like biased or wrong outputs. AI security is about stopping deliberate attacks and misuse. They overlap, because a successful attack can make a system unsafe.

Is AI governance the same as AI compliance? Not quite. Compliance means meeting specific legal requirements, such as the EU AI Act. Governance is the wider system of ownership, policies and records that makes compliance possible and keeps it going.

Does a small business need AI governance? If your team uses AI tools, yes, in proportion to how you use them. For most small businesses that starts with three things: knowing which tools you use, having a simple AI use policy, and naming who's responsible.

Who is responsible for AI safety, security and governance in a company? Safety usually sits with product and AI teams, security with IT, and governance with leadership. But governance is what makes sure the other two have an owner at all.

Is ISO 42001 required by law? No. ISO/IEC 42001 is a voluntary standard. It can help you organise and evidence how you manage AI, which supports legal compliance, but it's not a legal requirement in itself.

Not sure where your business stands?

Plenty of businesses are already using AI tools without a clear picture of which ones are in use, which risks matter, and what the law expects from them.

That's what our free consultation is for. In one call we'll look at how your business uses AI, show you where safety, security and governance apply to you, and tell you what's worth doing first. No jargon, no obligation.

This article is general information, not legal advice.

Sources


Comments


bottom of page