Indirect Prompt Injection
An attack in which malicious instructions reach a model through content it consumes rather than through the user's own prompt -- hidden in a web page, PDF, email, code comment, calendar invite, ticket or retrieved document. Because a language model does not reliably distinguish data from instructions, retrieved text can redirect its behaviour, exfiltrate context, or trigger tool calls. It is a distinct entry from direct prompt injection in the OWASP Top 10 for LLM Applications and is covered in CLLMSP, OSAI and AI Security Fundamentals.
Why It Matters
In practice this is the injection variant that matters most for real deployments, because the victim never sees the payload: a user asks an assistant to summarise a document and the document instructs the assistant to search their mailbox and post the results to an external URL. It is also why retrieval-augmented generation and tool use raise the stakes -- the more sources an agent reads and the more actions it can take, the more entry points exist. No prompt-level defence is complete, since instructing a model to ignore instructions is itself just text, so the durable controls sit around the model: treat all retrieved content as untrusted input, enforce authorisation on tool calls independently of the model, constrain outbound network destinations to prevent exfiltration channels, require human approval for consequential actions, and isolate high-privilege capability from any context that ingests untrusted data. Output encoding matters too, since model output rendered into a page or shell is a downstream injection risk in its own right. On exams, expect questions distinguishing it from direct injection and identifying which controls survive it.
Related AI Security terms
Prompt Injection
An attack against large language model (LLM) applications in which crafted input manipulates the model into ignoring its original instructions or system prompt and performing attacker-controlled actions. Direct prompt injection embeds malicious instructions in user input ("ignore previous instructions and..."), while indirect prompt injection hides instructions in external content the model ingests (web pages, documents, emails) during retrieval or tool use. It ranks as the #1 risk in the OWASP Top 10 for LLM Applications. Prompt injection is a core topic in AI security and governance certifications such as AIGP, AICP, and AAISM.
Jailbreaking (LLM)
Techniques that bypass an AI model's safety guardrails and content policies to elicit prohibited outputs such as instructions for weapons, malware, or disallowed content. Common methods include role-play framing ("act as an unrestricted assistant"), obfuscation and encoding, many-shot priming, and adversarial suffixes discovered through optimization. Jailbreaking differs from prompt injection: jailbreaking targets the model's safety alignment, whereas prompt injection hijacks an application's surrounding instructions. It is central to red-teaming generative AI and appears in AICP, AIGP, and AAISM study domains.
Adversarial Examples
Inputs deliberately perturbed with small, often human-imperceptible changes that cause a machine learning model to misclassify them — for example altering a few pixels so an image classifier reads a stop sign as a speed-limit sign, or crafting audio that a voice assistant transcribes as a hidden command. Adversarial machine learning is the broader field studying such evasion attacks alongside poisoning and extraction across the ML lifecycle. NIST formalizes the taxonomy in NIST AI 100-2. Covered in AAISM, AICP, and AIGP.
Data Poisoning
An attack in which adversaries inject malicious or mislabeled data into a model's training set to degrade performance, cause targeted misclassifications, or implant a backdoor that activates on a specific trigger. Poisoning can target the pre-training corpus, fine-tuning data, or a retrieval (RAG) knowledge base. Because modern models train on large, often web-scraped datasets, even a small fraction of poisoned samples can have outsized effects. It appears in the OWASP Top 10 for LLM Applications and NIST AI 100-2, and is tested in AAISM, AICP, and AIGP.
Model Inversion Attack
A privacy attack that reconstructs sensitive training data, or attributes of it, by repeatedly querying a model and analyzing its outputs — for example recovering recognizable face images from a facial-recognition model or inferring private attributes of individuals in the training set. Model inversion undermines the confidentiality of the data a model was trained on and can breach privacy regulations such as GDPR. It is a key privacy risk in AI governance and is covered in AICP, AIGP, and AAISM.
Membership Inference Attack
A privacy attack that determines whether a specific data record was part of a model's training set, by exploiting differences in the model's confidence or behavior on data it has seen versus unseen data. It can reveal, for instance, that a particular person's medical record was used to train a model — a confidentiality breach in its own right. Membership inference is closely related to model inversion and is relevant to AICP, AIGP, and privacy-focused AI governance.