What is blast radius in AI, and how do you keep it small?

A question our customers keep asking us is: "What does risk actually look like?" They've heard about AI being tricked, making mistakes or getting hacked. What they want to know is how bad it could get for their business if it happened.
The most useful way we've found to answer that is with an idea security teams have used for years, called blast radius.
What does blast radius mean in AI?
Blast radius is how much damage an AI system can do when something goes wrong: how much data it can reach, what it can change, and how many people or systems are affected before someone stops it.
A chatbot that can only answer questions from your publicly available information has a small blast radius. If it gets tricked, the worst it can usually do is give a wrong or embarrassing answer, which can still cause problems but is limited. An AI assistant that can read every inbox, update your customer records and send emails on your behalf has a bigger one. If that gets tricked, it can potentially leak, change or delete information before anyone notices.
Both could be running on exactly the same AI model. The difference in risk comes from what each one is connected to and what it has been allowed to do.
Why AI can make blast radius bigger
AI tools now take actions, not just give answers. Many AI systems are connected to other tools so they can call functions and work with other systems on someone's behalf: sending emails, updating records, running code. Each new connection adds to what could go wrong.
They're often given more access than they need. OWASP, the security community behind a widely used list of AI application risks, puts this down to too much functionality, too many permissions and too much autonomy. Their example is an assistant that only needs to read and summarise emails, but has been connected with a tool that can also send them. It was never meant to send anything, but it has the ability to.
You may not be able to fully stop them being tricked. In December 2025, the UK's National Cyber Security Centre warned that prompt injection (hiding instructions in content an AI reads) may never be fully fixed, because AI models can't reliably tell the difference between instructions and data. Its advice was to focus on reducing the impact instead, which is exactly what managing blast radius means.
They work quickly. An AI agent can carry out many actions in a few seconds, so by the time someone notices a problem, a lot may already have happened.
Two real examples
When the AI itself caused the damage
In July 2025, SaaStr founder Jason Lemkin was building an app with Replit's AI coding agent as a public experiment. During a code freeze, with instructions not to change anything without asking, the agent deleted his production database, which held records on more than 1,200 executives and more than 1,190 companies.
There was no attacker. The agent simply had access to the live database, so when it got things wrong, it damaged the live database. The agent also told him the data couldn't be restored, which turned out to be wrong, and he got it back. Replit's CEO called the deletion unacceptable, and the company's fix was separating test and live databases automatically so the agent couldn't reach live data while building. That change limited what the agent could reach, which is what managing blast radius looks like in practice.
When an attacker used an AI tool's access
In August 2025, attackers used stolen access tokens from Drift, an AI chatbot that Salesloft sells to sales and marketing teams, to pull data from customers' Salesforce systems. Google estimated that more than 700 organisations may have been affected, and companies including Cloudflare, Palo Alto Networks and Zscaler confirmed they were hit.
The chatbot itself wasn't tricked. Investigators found the attackers had got into Salesloft's GitHub account as early as March 2025, and Drift's connections to its customers' Salesforce accounts became the way in. Drift was connected to the customer data of hundreds of companies, so a breach at one supplier affected all of them. Google later advised anyone using Drift to treat all its connection tokens as compromised and replace them.
This could have happened just as well with a non-ai system, but it illustrates why you should keep the permissions of any credentials you create as restricted as possible. If they ever fall into the wrong hands you want to minimise the damage they can do.
Is blast radius different from other AI cyber security threats?
It's a fair question, and the short answer is that blast radius isn't a threat in itself. Prompt injection, data poisoning and stolen credentials are ways an AI system can be attacked or go wrong. Blast radius describes how much damage follows when that happens.
Threat | Blast radius | |
The question it answers | How could this go wrong? | How bad will it be when it does? |
Examples | Prompt injection, data poisoning, stolen tokens, AI mistakes | Data leaked, records deleted, number of customers affected |
How you deal with it | Make it harder to happen | Limit what the AI can reach and do |
IT teams have limited blast radius for years by giving staff only the access they need. There are two reasons it matters more with AI.
First, the AI can cause the damage on its own, with no attacker involved, as the Replit case showed. Second, for weaknesses like prompt injection there's no reliable fix yet, so for now, limiting the damage is the main line of defence.
How to manage the blast radius of your AI tools
Start with one question for every AI tool you use: if this did the worst thing it's technically able to do, what would happen? Then work through these.
Know what each tool can reach. List your AI tools and, for each one, the data, systems and accounts it's connected to. You may find connections nobody remembers setting up.
Give it the least access that works. If a tool only needs to read, don't let it write or delete. OWASP's example is connecting an email assistant with read-only access when all it does is summarise.
Remove what it doesn't use. Old plugins, trial integrations and extra features all widen the blast radius for no benefit.
Keep a person in the loop for big actions. Sending money, deleting records, emailing customers or changing live systems should need a human to approve first.
Keep testing away from live systems. The Replit fix in one line: an AI that's building or experimenting shouldn't be able to touch real customer data.
Slow it down. Limits on how many actions a tool can take in a given time give you a chance to spot problems before the damage grows.
Log and watch what it does. The NCSC recommends live monitoring, including of failed tool calls, which can be a sign someone is testing their way in.
Treat third-party AI connections as part of your risk. When you connect an outside AI tool to your systems, a security problem at that supplier can quickly become your problem too. Know which connections exist and how to cut them off quickly, as Drift customers had to.
Decide who pulls the plug. Name the person who can switch a tool off, and make sure they know how, before you need them to.
Blast radius is a governance decision
The technical fixes above are only part of the answer. Very often, nobody can say who decided what an AI tool should have access to. Someone set it up, clicked "allow" on the permissions, and moved on.
Every AI tool's blast radius was decided by someone, even if they didn't realise they were making a decision. Good governance means those decisions are made on purpose, by the right person, and written down. In practice, that means being able to answer these questions for each AI tool you use:
Who approved it, and who owns it now?
What is it connected to, and why does it need that access?
When was that access last reviewed?
Who can switch it off, and do they know how?
If you can answer those, you're in a good position, both for managing the risk and for showing a regulator, customer or auditor that you have it under control. This is the part we help with. We're not there to configure your systems, but we help you make and record these decisions so the technical side has something clear to follow.
Where UK and European law and standards come in
No law uses the phrase "blast radius", but several expect you to manage it.
EU AI Act. Article 15 requires high-risk AI systems to be resilient against attempts to manipulate them and to have appropriate cybersecurity. The Act also requires human oversight of high-risk systems. Following the Digital Omnibus, these obligations apply from 2 December 2027 for high-risk uses such as hiring and credit scoring, and from 2 August 2028 for AI built into regulated products. UK businesses can be caught if they sell AI into the EU or their AI's output is used there.
UK GDPR. If an AI tool leaks or exposes personal data, it's a personal data breach like any other. If it's likely to put people's rights and freedoms at risk, you must report it to the ICO without undue delay, and within 72 hours where feasible. Limiting what your AI tools can reach also limits how much personal data could be exposed if one of them is misused.
NCSC guidance. It isn't law, but the UK's National Cyber Security Centre is clear that organisations using AI should design for the risk of prompt injection and limit what a compromised AI can do.
ISO/IEC 42001. The international standard for AI management systems asks you to assess AI risks and put controls in place. Mapping what each AI tool can reach, and limiting it, is a practical way to show you've done that.
Frequently asked questions
What is blast radius in AI? It's how much damage an AI system can do when something goes wrong: the data it can reach, what it can change, and how many people or systems are affected before it's stopped.
Is blast radius the same as prompt injection? No. Prompt injection is one way an AI can be attacked. Blast radius is how much damage follows. A prompt injection against a tool with very limited access does little harm. The same attack against a tool with wide access can be serious.
How do I reduce the blast radius of an AI agent? Give it only the access it needs, prefer read-only, remove tools it doesn't use, require human approval for high-impact actions, keep it away from live systems while testing, and log what it does.
Does this apply to everyday AI assistants, not just custom agents? Yes. Once an assistant is connected to your email, files, calendar or CRM, it can reach whatever those connections allow. Each extra connection increases what it could expose or change.
Do small businesses need to worry about this? If you use AI tools connected to customer data, finance or live systems, yes. You don't need a security team to start. Listing what each tool can reach and removing what it doesn't need covers a lot of ground.
Want to know how far your AI could reach?
If you use AI tools at work, try answering two questions: what is each one connected to, and what would happen if it went wrong? If either is hard to answer, that's where to start.
That's what our free consultation is for. We'll go through the AI tools your business uses, talk through what each one is connected to and who's responsible for it, and give you the questions to ask your IT provider or supplier. You'll come away with a clear picture of where your risk sits and what to do first.

This article is general information, not legal advice.
Sources
LLM06:2025 Excessive Agency, OWASP GenAI Security Project
NCSC warns of confusion over true nature of AI prompt injection, Computer Weekly, December 2025
Prompt injection attacks might 'never be properly mitigated', UK NCSC warns, TechRadar, December 2025
An AI-powered coding tool wiped out a software company's database, then apologized for a 'catastrophic failure on my part', Fortune, 23 July 2025
Salesloft Takes Drift Offline After OAuth Token Theft Hits Hundreds of Organizations, The Hacker News, 3 September 2025
News brief: Salesloft Drift breach update and timeline, TechTarget, 2025
Hackers Steal OAuth Tokens via Salesloft Drift Integrations in Massive Salesforce Data Theft, CPO Magazine, 2025
EU AI Act, Article 2: Scope, European Commission AI Act Service Desk
EU AI Act, Article 15: Accuracy, robustness and cybersecurity, European Commission AI Act Service Desk
Digital Omnibus on AI, AI Act Explorer (Future of Life Institute)
Personal data breaches: a guide, Information Commissioner's Office



Comments