AI vulnerabilities
AI has weak spots, like any technology. These are the ten that security experts rank highest, explained in plain English, with what you can do about each one.
- 01
Prompt injection
Hidden instructions that hijack the AI
Someone slips instructions into content the AI reads, and the AI follows them instead of you. See .
- What it looks like
- A web page contains invisible text telling your AI assistant to ignore your request and send your details somewhere else.
- How to stay safe
- Be cautious about letting an AI with access to your accounts read unknown websites, emails or files. Builders should treat all outside content as untrusted.
- 02
Sensitive information disclosure
The AI reveals things it shouldn't
Private data comes out in an answer, either because it was in the or because someone fed it in.
- What it looks like
- A support chatbot repeats another customer's details, or a colleague pastes a confidential contract into a public chatbot.
- How to stay safe
- Never paste passwords, client data or anything confidential into a public AI tool. Check your tool's data and privacy settings.
- 03
Supply chain risks
Dodgy parts from someone else
AI apps are assembled from other people's , plugins and datasets. If one of those is tampered with, everything built on it is affected.
- What it looks like
- A free model downloaded from the internet has been quietly altered to behave badly in certain situations.
- How to stay safe
- Use models, plugins and extensions from reputable sources, and keep them updated, just as you would with phone apps.
- 04
Data and model poisoning
Teaching the AI bad lessons on purpose
An attacker corrupts the material a model learns from, so it picks up false facts, or a hidden trigger.
- What it looks like
- Fake reviews and planted web pages are created so that an AI learns to recommend a scam product.
- How to stay safe
- Builders need to vet where training data comes from. Everyone else: be wary of unknown models and cross-check surprising claims.
- 05
Improper output handling
Trusting the AI's answer blindly
An app takes whatever the AI produces and runs it or displays it without checking, so a bad answer becomes a real action.
- What it looks like
- AI-written code or a link is executed automatically and it turns out to be harmful.
- How to stay safe
- Read AI-generated code, commands and links before you run or click them. Builders should check AI output like any untrusted input.
- 06
Excessive agency
Giving the AI too much power
An is allowed to do more than the job needs, such as deleting files, spending money or sending emails, so one mistake does real damage.
- What it looks like
- An assistant asked to tidy an inbox has permission to delete everything, and does.
- How to stay safe
- Give AI tools the minimum access they need, and keep a for anything important or irreversible.
- 07
System prompt leakage
The hidden instructions get out
The behind-the-scenes is revealed, and sometimes it contains secrets that should never have been there.
- What it looks like
- A user coaxes a company chatbot into printing its hidden instructions, which include an internal password.
- How to stay safe
- Builders should never put passwords, keys or confidential rules in a prompt. Assume anything in a prompt can be seen.
- 08
Vector and embedding weaknesses
Flaws in the AI's filing cabinet
Systems that let an AI look things up ( and ) can expose documents to the wrong people or be fed false ones.
- What it looks like
- A company knowledge assistant answers a junior employee's question using a confidential board document.
- How to stay safe
- Apply the same access permissions to the AI's knowledge base as to the original files, and control who can add content to it.
- 09
Misinformation
Confident answers that are wrong
The AI states something false and people act on it. This is turning into real-world harm.
- What it looks like
- A chatbot invents a legal case, a medical dose or a software package that doesn't exist, and someone relies on it.
- How to stay safe
- Verify anything important with a trusted source, ask the AI where its information came from, and don't use it as your only adviser.
- 10
Unbounded consumption
Running up the meter
Someone floods an AI service with requests, which can knock it offline or create an enormous bill, since AI is charged by usage.
- What it looks like
- A public chatbot with no limits is hammered by automated traffic overnight and the owner wakes up to a huge invoice.
- How to stay safe
- Builders should set usage limits, spending alerts and rate limits from day one.
Based on the OWASP Top 10 for Large Language Model Applications, the security industry's reference list. The explanations are our own.
One email a fortnight. Five stories. No jargon.
Every other Friday: the AI news that matters, explained in plain English, plus one new word to learn. Unsubscribe any time.