Prompt injection is when text an AI model reads, whether typed by a user or hidden in a document, email or web page, overrides the instructions its developers gave it. It is serious because there is no complete fix: the NCSC says it may never be fully mitigated. How much harm it causes depends on what the AI can access and do.
Prompt injection sits at the top of the OWASP Top 10 for LLM Applications, in both the 2025 and 2026 editions. For most organisations, the useful question is not whether an AI application can be prompt-injected. It almost certainly can. The question is what happens when it is.
Executive Summary & Key Takeaways
- Prompt injection is number one in the OWASP Top 10 for LLM Applications 2026, published in August 2026, keeping the position it held in 2025.
- It comes in two forms: direct, where a user types the attack, and indirect, where instructions are hidden in content the model reads. Indirect injection is the bigger business risk because the attacker needs no access to your systems.
- The NCSC warns that, unlike SQL injection, prompt injection may never be fully mitigated, because language models do not separate instructions from data.
- Severity scales with access. A chatbot with no data or tools is a nuisance; an assistant that can read your inbox and send email is a potential data breach.
- Controls that limit what a manipulated model can do matter more than filters that try to spot attacks.
How prompt injection works
An AI application usually combines its developer’s instructions (often called the system prompt) with other text: the user’s question, retrieved documents, emails, web pages or the output of tools. The model receives all of it as one stream and predicts what comes next. If any part of that stream contains instructions, the model may follow them.
Direct prompt injection is the familiar version: a user types something like “ignore your previous instructions” to make a chatbot break its rules. Jailbreaking, where the aim is to make a model disregard its safety protections entirely, is a form of prompt injection (OWASP, retrieved 3 November 2026).
Indirect prompt injection is more dangerous for businesses. An attacker plants instructions in content your AI will later read. The NCSC gives the example of a CV containing hidden text telling a screening system to approve the candidate for interview (NCSC, retrieved 3 November 2026). OWASP describes a web page with hidden instructions that cause an assistant summarising it to send the user’s private conversation to an attacker-controlled address (OWASP, retrieved 3 November 2026). The user never sees the attack, and the attacker never touches your systems.
MITRE ATLAS, the knowledge base of attacker techniques against AI systems, records LLM prompt injection as a technique with direct, indirect and triggered variants (MITRE, retrieved 3 November 2026).
Why it cannot simply be patched
Security teams sometimes treat prompt injection as a cousin of SQL injection, which was largely tamed by keeping code and data apart through parameterised queries. The NCSC argues that this comparison is dangerous. Inside a large language model there is no distinction between data and instructions, only ever the “next token”, so prompt injection may never be mitigated in the way SQL injection can be (NCSC, “Prompt injection is not SQL injection (it may be worse)”, 8 December 2025, retrieved 3 November 2026).
The NCSC instead describes language models as an “inherently confusable deputy”: a component that can be talked into using its privileges on an attacker’s behalf. OWASP reaches a similar conclusion, saying it is unclear whether fool-proof prevention methods exist (OWASP, retrieved 3 November 2026).
The practical consequence is that prompt injection is a residual risk to be managed through design, not a bug to be fixed once. The NCSC advises wariness of any product that claims to “stop” it.
How serious is it for your organisation?
Severity depends on three things: what untrusted content the model reads, what data it can reach and what actions it can take. The OWASP 2026 list cites the “lethal trifecta”: a system that can access private data, ingest untrusted content and communicate externally has the conditions for high-impact exploitation, and removing any one of the three removes those conditions (OWASP, retrieved 3 November 2026).
| model | What a successful injection could do | Typical severity |
|---|---|---|
| Public chatbot answering from public information, with no tools | Misleading or off-brand answers, reputational embarrassment | Low to moderate |
| Internal assistant over company documents | Return manipulated answers, reveal its instructions, surface documents a user should not see | Moderate to high |
| Assistant connected to email, calendar or files | Read data and send it to an attacker, sometimes with no user interaction | High |
| AI agent that takes actions in business systems | Change records, approve transactions, run code or call other tools on an attacker’s instructions | High to critical |
The last row is where prompt injection meets excessive agency, which rose to third place in the 2026 OWASP list. We cover that in the security risks of AI agents.
What reduces the risk
Because injection cannot be reliably blocked, the most effective controls limit its consequences. At a buyer level, these are the questions to ask of any AI system you build or buy.
- Does it need that access? The UK Code of Practice for the Cyber Security of AI requires that permissions an AI system holds on other systems are only those needed for its function, and are risk assessed (provision 2.6) (GOV.UK, retrieved 3 November 2026).
- Do its privileges drop when it reads untrusted content? The NCSC highlights a simple design rule: when a model processes information from a party, its privileges should drop to that party’s level. A model reading emails from strangers should not hold privileged tools (NCSC, retrieved 3 November 2026).
- Are high-impact actions checked by something other than the model? Deterministic controls, such as fixed approval rules or human sign-off, constrain what the system can do regardless of what the model has been told.
- Is untrusted content kept separate and labelled? Marking external content as data makes injection harder, though not impossible.
- Is it logged and monitored? The NCSC recommends logging enough to spot an attacker refining their attack, potentially including full model inputs and outputs, tool use and API calls.
- Has it been tested by people trying to break it? Testing should include indirect routes, such as uploaded files, retrieved documents and tool outputs, not just the chat box. This is a core part of AI red teaming.
For the wider set of controls around AI systems, see our guide to what AI security is.
Common mistakes
- Treating prompt injection as solved because a filter blocks obvious phrases like “ignore previous instructions”.
- Testing only the chat interface and ignoring the documents, emails and web content the model reads.
- Giving an assistant broad access to data or tools for convenience, before assessing what an attacker could do with it.
- Relying on the system prompt to enforce security rules, when the system prompt is the very thing being overridden.
- Assuming a third-party AI product has dealt with the risk, without asking how.
Frequently asked questions
Is jailbreaking the same as prompt injection?
Jailbreaking is one type of prompt injection. OWASP describes it as the form where an attacker’s input makes a model disregard its safety protocols entirely. Prompt injection is broader and includes attacks that leave safety rules intact but redirect the model, for example to leak data or misuse a tool.
Can a guardrail product or AI firewall solve prompt injection?
It can reduce the risk, not remove it. The NCSC warns that blocking known attack phrases fails because there are endless ways to rephrase an attack, and advises caution with any product that claims to stop prompt injection. Use such tools as one layer, alongside limits on what the system can access and do.
Will a newer or better model fix it?
Not on its own. Training models to prioritise developer instructions over instruction-like text in data is an active research area, but the NCSC notes that these approaches try to impose a distinction the technology does not inherently make. OWASP also says that retrieval augmented generation and fine-tuning do not fully mitigate prompt injection.
Should prompt injection be on our risk register?
Yes, for any AI system that reads untrusted content. The UK AI code of practice expects threats to AI systems to be analysed and managed, and recognises that a higher level of risk will remain in AI systems despite controls. Recording prompt injection as a residual risk, with an owner and the controls that limit its impact, is the honest position.
Getting started
List the AI systems you run or buy, what content each one reads and what each one can reach or do. That shows you where prompt injection would hurt most. Our AI and ML security team reviews and tests AI applications against current OWASP and MITRE guidance, or talk to our team about a specific system.

Jared Henderson
Writing about cyber security and technology at Falx.





