What is AI safety, and how is it different from AI security

AI safety is about stopping AI tools from causing harm, whether by accident, through misuse or through behavior nobody intended. AI security is about protecting those tools, your data and the people who rely on them from attackers. They overlap, but they ask different questions, and a small organization needs to think about both.

What AI safety means

When people ask “what is AI safety,” they usually get one of 2 very different answers. One is about the tools you use today. The other is about powerful systems that might exist in the future.

For AI safety in everyday use, the problems are practical:

Wrong answers stated with confidence. A chatbot can invent a fact, a quote or a legal rule and present it as true. This is often called a hallucination.
Harmful or biased output. A tool can write something offensive, unfair or unsuitable for the people you serve.
Trusting the output too much. Staff can stop checking because the answers usually look right. This is called overreliance.

The longer-term side of AI safety asks whether more capable systems might act against what their makers intend. This is usually called alignment. It is a real field of work, and it is what most news coverage focuses on, but it is not the part that will affect your volunteer schedule or your grant report next week.

Put simply, AI safety is about harm the tool could do, even when nobody is attacking it.

How AI safety differs from AI security

AI security is about attackers. Someone wants to steal your data, trick the tool into doing something it should not, or break the software it runs on. Common examples include:

Prompt injection: hidden instructions in a document, email or web page that take over what the AI tool does.
Data poisoning: tampering with the information a model learns from, so it behaves badly later.
Data leaks: private information escaping through the tool, its settings or its vendor.
Ordinary software flaws: AI products are software, and they have bugs like any other software.

Here is how the 2 fields compare for a small organization.

	AI safety	AI security
Main worry	The tool causes harm by mistake or through misuse	Someone attacks the tool or your data
Typical problem	A false answer, a biased summary, staff trusting output blindly	A leaked donor list, a hijacked assistant, an unpatched app
Questions to ask	Can we trust this output? Could it harm someone we serve? Who checks it?	Who could get our data? What could an attacker make the tool do? Is the software up to date?
Who usually acts	The people who use the output and decide what to do with it	Whoever manages your accounts, settings and software

Where they overlap

The line between the 2 is not always clean. A few examples:

Prompt injection is a security attack, but the result can be a safety failure. A hijacked assistant might send a client false advice or harmful content.
A jailbreak is a trick that gets a tool to ignore its safety rules. It is an attack used to defeat safety protections.
A deepfake used for fraud is misuse of AI. A fake voice or video of your director asking for an urgent payment is a safety problem caused by someone acting in bad faith.

You do not need to file every problem in the right box. What matters is that someone in your organization is asking both sets of questions.

Everyday problems versus the news debate

Most headlines about AI safety focus on the future: whether advanced systems could become hard to control. That debate is worth following, but it can make AI safety feel distant or abstract.

For a small organization, the safety problems you will meet are closer to home. Here are some illustrations of the kind of thing that happens:

False output. A staff member asks a chatbot to summarize a funding rule, and it describes a rule that does not exist. The summary goes into a grant application.
Harmful or biased output. A tool drafts outreach messages that use stereotypes about the community you serve, or writes a client letter in a tone that would upset someone in a hard situation.
Overreliance. A volunteer uses AI to answer questions from the public and stops checking the answers after the first week, because they all looked fine.
Misuse such as deepfake fraud. Your finance person gets a voice message that sounds exactly like your director, asking for a transfer before the end of the day.

None of these needs a futuristic AI. They happen with the tools available now, and they are the ones worth planning for first.

What a small organization can do

You do not need a technical team to reduce most of these risks. These steps work for an organization of any size.

Have a person review AI output before it reaches people. Anything that goes to a client, donor, funder or the public should be read and checked by a staff member first. This is the single most useful safety habit.
Do not use AI alone for decisions about individuals. Decisions about who gets a service, a place, a grant or a job should be made by people. AI can help draft or sort, but a person decides and can explain why.
Test tools on your own typical tasks. Before adopting a tool, try it on the real work you would use it for. Give it a few examples where you already know the right answer and see how often it gets them wrong.
Teach staff the limits. Explain that AI tools can be confidently wrong, can repeat bias, and should be checked. A short conversation at a team meeting goes a long way.
Verify unusual requests through a second channel. If a message asks for money, passwords or sensitive data, especially in a hurry, confirm it by calling the person on a number you already have. This is your main defense against deepfake fraud.
Keep the security basics in place. Agree on which AI tools staff may use, check their privacy and sharing settings, and keep the software up to date.

Matching steps to problems

False output: steps 1, 3 and 4.
Harmful or biased output: steps 1, 2 and 3.
Overreliance: steps 1 and 4, and repeat the message over time, because the habit of checking fades.
Deepfake fraud: step 5, and make sure everyone who handles money knows the rule.
Security problems: step 6. For more on what happens to the information you type into AI tools, see the course lesson on what AI tools do with your data, and for attacks on chatbots, read LLM security basics.

Where frameworks fit

You may hear about formal frameworks for managing AI risk. The best known in the United States is the AI Risk Management Framework from NIST, the US National Institute of Standards and Technology. Version 1.0 was released on January 26, 2023.

The framework is voluntary. Its aim is to help organizations build trustworthiness into how AI is designed, developed, used and evaluated, and it considers risks to individuals, to organizations and to society. It is organized around 4 functions: Govern, Map, Measure and Manage.

For a small organization, you do not need to adopt the whole framework. It helps most as a checklist of ideas:

Govern: decide who is responsible for AI use and write down your rules.
Map: list where you use AI and who could be affected.
Measure: check how well the tools work on your tasks.
Manage: act on what you find, and review it from time to time.

The 6 steps above already cover a good part of this. A framework becomes more useful as your AI use grows, or when a funder or partner asks how you manage AI risk.

How Securaity covers both

This site tracks both sides. We publish advisories when we find security problems in AI software and services, with a plain rating of how serious they are and what to do. For safe everyday use, our tool profiles, course and guides explain what popular tools do and how to use them carefully.

Frequently asked questions

Is AI safety only about future superintelligent AI?

No. The future debate gets the most attention, but AI safety also covers the problems you meet today: false answers, biased or harmful output, and people trusting tools too much. For a small organization, these everyday problems are where your effort pays off.

Who is responsible for AI safety, the vendor or us?

Both. The vendor is responsible for how the tool is built and what protections it includes. You are responsible for how your organization uses it, including what you check, what decisions you let it influence, and what reaches the people you serve.

Can AI output be biased even if we do not use it for decisions?

Yes. Bias can show up in drafted emails, summaries, images or outreach messages, and it can still hurt or put off the people you work with. Having a person review output before it goes out catches most of it.

How do we start?

Pick one person to own the topic, list the AI tools your staff already use, and agree on 2 rules: a person reviews output before it reaches anyone, and unusual requests for money or data get checked by phone. Then test your main tools on a few real tasks and build from there.