# Prompt injection: meaning and business risks | Aeptus - Kind: Detail - URL: /glossary/prompt-injection - Canonical: https://www.aeptus.com/glossary/prompt-injection/ - HTML lang: en-GB - H1: Prompt injection - Description: Prompt injection uses untrusted instructions to make an AI system depart from its intended task or misuse its access. - Structured data: BreadcrumbList, DefinedTerm, ListItem, WebPage - Internal links: 81 - Image coverage: 5/13 images with alt text - Summary: Prompt injection: meaning and business risks | Aeptus. Prompt injection uses untrusted instructions to make an AI system depart from its intended task or misuse its access.. Prompt injection - Related links: product-connect-and-prepare-data, product-collect-and-keep-evidence-current, product-report-with-confidence, work-with-ai-agents, automation - Outline: Section: Latest articles Canadian report maps security work in small businesses A review of public data looks at security staffin | Section: Editors’ picks Unauthorized radio signals stopped trains in Poland in 2023 In August 2023, unauthorized radio transmissi | Section: Explore Make security decisions 1 Prepare, respond and recover 2 Protect everyday business 2 Prove security and earn tru ## Section Latest articles Canadian report maps security work in small businesses A review of public data looks at security staffing, written policies and updates among Canadian businesses with 10 to 49 employees. INTERCEPT hosts free cybersecurity webinar for small teams The 18 November online event will cover phishing, multi-factor authentication and a cybersecurity learning platform. A third-party software vulnerability led to a data breach at Frontline Education Frontline has begun notifying people affected by an August incident involving two school administration products. The impact varies between districts and individuals. ## Section Editors’ picks Unauthorized radio signals stopped trains in Poland in 2023 In August 2023, unauthorized radio transmissions triggered Poland’s railway stop function. Reporting said the receiver did not verify who sent the command; no one was injured. A cyberattack disrupted refrigeration at hundreds of French supermarkets An attack on 12 October 2026 changed refrigeration settings and led stores to discard fresh and frozen food. CERT-FR named several affected controller families, warned that activity might continue and did not identify how access began. A secret is not rotated until the old credential stops working Railway’s preliminary report says some environment variables in at most 1.85% of workspaces may have been compromised, but it could not identify which specific variables were affected. Prove a rotation is complete by checking each consumer and confirming the old credential is revoked. ## Section Explore Make security decisions 1 Prepare, respond and recover 2 Protect everyday business 2 Prove security and earn trust 1 Understand threats and change 9 ## Section Topics ## LearnEditorialMobileSection Latest articles Canadian report maps security work in small businesses A review of public data looks at security staffing, written policies and updates among Canadian businesses with 10 to 49 employees. INTERCEPT hosts free cybersecurity webinar for small teams The 18 November online event will cover phishing, multi-factor authentication and a cybersecurity learning platform. A third-party software vulnerability led to a data breach at Frontline Education Frontline has begun notifying people affected by an August incident involving two school administration products. The impact varies between districts and individuals. View all articles → ## LearnEditorialMobileSection How-to Guides Answer security questionnaires without chaos A step-by-step way to respond faster without improvising every answer from scratch. View all guides → ## LearnEditorialMobileSection Editors’ picks Unauthorized radio signals stopped trains in Poland in 2023 In August 2023, unauthorized radio transmissions triggered Poland’s railway stop function. Reporting said the receiver did not verify who sent the command; no one was injured. A cyberattack disrupted refrigeration at hundreds of French supermarkets An attack on 12 October 2026 changed refrigeration settings and led stores to discard fresh and frozen food. CERT-FR named several affected controller families, warned that activity might continue and did not identify how access began. A secret is not rotated until the old credential stops working Railway’s preliminary report says some environment variables in at most 1.85% of workspaces may have been compromised, but it could not identify which specific variables were affected. Prove a rotation is complete by checking each consumer and confirming the old credential is revoked. ## LearnEditorialMobileSection Explore Article categories Make security decisions 1 Prepare, respond and recover 2 Protect everyday business 2 Prove security and earn trust 1 Understand threats and change 9 Topics ## GlossaryTermProof Glossary Prompt injection Jump to Definition References Definition Prompt injection uses untrusted instructions to make an AI system depart from its intended task or misuse its access. Instructions can arrive directly from a person or indirectly inside a document, email or web page the AI reads. Consequences depend on the system's permissions. A customer may include “Prompt injection” in a supplier security questionnaire or spreadsheet. They want to understand whether this topic affects the service you provide and what safeguards or evidence support your answer. Prompt injection is an attempt to make an AI system follow untrusted instructions that conflict with its intended task. The instructions may come from a person using the system directly or be embedded in material it reads, such as a document or website. The risk grows when an AI assistant can retrieve private information, send messages or change records. Text that should be treated as content can instead influence an action. Even without connected tools, injection can distort a summary or answer. It is distinct from an ordinary factual mistake in the model's response. Hypothetical example: a small business asks an assistant to summarise supplier proposals. One proposal contains hidden instructions intended to make the assistant favour that supplier or disclose information from another document. The document should not have authority over the assistant's task. Give AI services only the information and permissions needed for their work, and require suitable checks before sensitive actions. Treat text from documents, web pages and user input as untrusted data. Limit what an AI system can access or change, test whether embedded instructions alter its behaviour, and require human approval before sensitive actions. Filtering alone is not a complete safeguard. Filtering prompts alone is not a complete guarantee. The business also needs to decide which decisions require a person, especially when the assistant can act on external systems. Sources: OWASP GenAI Security Project: prompt injection. References OWASP GenAI Security Project: prompt injection Wikipedia: Prompt injection Page progress 0% You’re here On this page Definition References How this term connects Related terms Social engineering Social engineering is manipulation that persuades a person to reveal information, grant access or take an unsafe action. Data exfiltration Data exfiltration is the unauthorised transfer of information out of a system or business environment. Privileged access Privileged access grants elevated permissions to manage systems, change security settings or reach sensitive information. ## GlossaryDefinitionSummary Definition Prompt injection uses untrusted instructions to make an AI system depart from its intended task or misuse its access. Instructions can arrive directly from a person or indirectly inside a document, email or web page the AI reads. Consequences depend on the system's permissions.