Crafted inputs manipulate an LLM into ignoring system instructions or taking unintended actions. Direct injections overwrite the system prompt; indirect ones hide instructions in external content the model later reads.
Examples
-
Direct Prompt Injection
A malicious user injects prompts that extract sensitive information or override safety rules.
-
Indirect Prompt Injection
Untrusted webpage or document content steers the model when it is summarized or retrieved.
-
Scam Through Plugins
Websites or plugins abuse tool-calling to run scams or exfiltrate data.
Prevention
-
Privilege Control
Limit what the LLM can access and apply role-based permissions to tools.
-
Human Approval
Require explicit user consent before privileged or irreversible actions.
-
Segregate Content
Keep untrusted retrieved content clearly separated from system and user prompts.
-
Trust Boundaries
Treat model output as untrusted and visually mark when content may be unreliable.
Attack scenarios
-
Chatbot Remote Execution
Injection leads to unauthorized tool use through a chatbot.
-
Email Deletion
Indirect injection causes destructive mailbox actions.
-
Exfiltration via Content
Embedded prompts push private data into attacker-controlled channels.
LLM02Insecure Output Handling
Downstream systems accept LLM output without validation. That can turn model text into XSS, SSRF, privilege escalation, or remote code execution when it is rendered or executed as code.
Examples
-
Remote Code Execution
Model output is passed to a shell or interpreter without sanitization.
-
Cross-Site Scripting (XSS)
Generated JavaScript or Markdown is rendered unsafely in a browser.
Prevention
-
Zero-Trust Output
Treat LLM output like untrusted user input - validate and sanitize before use.
-
OWASP ASVS Guidance
Apply input validation and encoding standards to model-generated content.
-
Output Encoding
Encode output for the target context (HTML, URL, SQL, shell) before render or exec.
Attack scenarios
-
Chatbot Shutdown
Unvalidated output triggers a plugin to shut down or misbehave.
-
Sensitive Data Capture
Generated payloads send secrets to an attacker-controlled endpoint.
-
Destructive Queries
Model-crafted SQL or admin commands wipe or alter data.
LLM03Training Data Poisoning
Tampered training or fine-tuning data introduces backdoors, bias, or unsafe behavior. Sources can include Common Crawl-style corpora, books, crowd-sourced sets, and vendor fine-tunes.
Examples
-
Malicious Data Injection
Falsified samples are inserted during training or fine-tuning.
-
Biased Training Outputs
The model inherits and amplifies tainted labels or narratives.
-
Content Injection
Attackers seed public corpora so later models learn harmful patterns.
Prevention
-
Dataset Provenance
Track origin, licensing, and integrity checks for every training set.
-
Filtering and Review
Scrub toxic, secret, and adversarial samples before training.
-
Canary and Eval Suites
Test for backdoors and bias regressions after every train or fine-tune.
Attack scenarios
-
Misleading Outputs
Poisoned models promote biased or harmful content.
-
Toxic Data Injection
Malicious contributors push hate or unsafe material into open datasets.
LLM04Model Denial of Service
Attackers force expensive inference (huge context, recursive tools, adversarial loops) so availability drops or cloud bills spike. LLM cost and latency make this easier than classic DoS.
Examples
-
High-Volume Queuing
Flood the model with many concurrent expensive requests.
-
Resource-Consuming Queries
Unusual prompts force heavy context or tool recursion.
-
Continuous Input Overflow
Oversized inputs exhaust context windows and workers.
Prevention
-
Input Validation
Cap length, reject pathological patterns, and filter abusive content.
-
Resource Caps
Limit tokens, tools, and concurrency per user or API key.
-
Rate Limits and Quotas
Throttle bursts and alert on anomalous spend or latency.
Attack scenarios
-
Resource Overuse
One tenant starves shared hosted capacity for others.
-
Request Amplification
A page or bot multiplies costly model calls through automation.
LLM05Supply Chain Vulnerabilities
Third-party models, datasets, plugins, and libraries can ship vulnerabilities or backdoors into your LLM stack - similar to classic software supply-chain risk, with model artifacts added.
Examples
-
Package Vulnerabilities
Outdated Python or JS deps used by the app or serving stack.
-
Vulnerable Models
Risky pre-trained checkpoints used for fine-tuning or inference.
-
Poisoned Data
Crowd-sourced or vendor datasets that were already compromised.
Prevention
-
Supplier Evaluation
Vet model and plugin vendors, SLAs, and update policies.
-
Plugin Testing
Only ship tested, pinned, least-privilege plugins and tools.
-
Dependency Hygiene
SBOM, SCA scanning, and pinned versions for serving and training code.
Attack scenarios
-
Library Exploitation
Known CVEs in ML or web libraries become RCE in your stack.
-
Malicious Plugin
A marketplace plugin exfiltrates prompts or credentials.
LLM06Sensitive Information Disclosure
Models may leak PII, secrets, or proprietary text from training data, RAG stores, or system prompts. Weak filtering turns chat into an unintended data-exfil channel.
Examples
-
Incomplete Filtering
Responses still contain secrets that simple denylists missed.
-
Overfitting / Memorization
The model reproduces sensitive strings seen in training.
Prevention
-
Data Sanitization
Scrub secrets and PII from training and retrieval corpora.
-
Input and Output Filters
Detect and block sensitive patterns on both sides of the model.
-
Least Data in Context
Only retrieve fields the task needs; never dump full records.
Attack scenarios
-
Unintentional Exposure
One user receives another user’s private context.
-
Filter Bypass
Adversarial phrasing extracts PII past naive guards.
LLM07Insecure Plugin Design
Plugins and tools expose powerful actions with weak parameter validation or authz. Once the model can call them, prompt injection becomes remote control of your systems.
Examples
-
Single Field Parameters
Plugins accept free-form strings instead of typed, separated fields.
-
Configuration Overrides
Config strings let callers widen privileges or change destinations.
Prevention
-
Parameter Control
Enforce schemas, types, allowlists, and a validation layer on every tool.
-
OWASP Guidance
Apply ASVS-style checks to plugin interfaces as if they were public APIs.
-
AuthZ per Action
Authorize each tool call against the end user, not the model identity alone.
Attack scenarios
-
URL Manipulation
Attackers inject destinations via manipulated tool arguments.
-
Recon and Exploitation
Over-broad tools become scanners or exploit runners.
Agents get more functions, permissions, or autonomy than the task needs. When something goes wrong - or gets injected - the blast radius is large.
Examples
-
Excessive Functionality
Agents expose unused tools that attackers can still invoke.
-
Excessive Permissions
Plugins run with admin-like scopes “just in case.”
Prevention
-
Limit Plugin Functions
Ship only the tools required for the product surface.
-
Plugin Scope Control
Restrict each tool to the minimum data and actions.
-
Human-in-the-Loop
High-impact steps need confirmation, not full autonomy.
Attack scenarios
-
Tricked Assistant
A malicious email or page steers an over-privileged personal assistant.
People or systems trust model output without review. Hallucinations become policy, insecure code gets shipped, and legal or security decisions rest on fluent nonsense.
Examples
-
Misleading Info
Teams accept incorrect claims because the answer sounds confident.
-
Insecure Code
Suggested snippets introduce vulnerabilities that nobody reviews.
Prevention
-
Monitor and Validate
Review outputs with consistency checks and domain experts where it matters.
-
Cross-Check Sources
Verify critical claims against trusted documentation and tests.
-
Safe Defaults
Do not auto-merge or auto-execute model-generated changes.
Attack scenarios
-
Disinfo Spread
Attackers exploit LLM-reliant publishers to amplify falsehoods.
-
Plagiarism / IP Risk
Unchecked generation creates compliance and copyright exposure.
Unauthorized access, copying, or extraction of proprietary models - via APIs, insider leak, or side channels - destroys competitive advantage and may expose embedded sensitive knowledge.
Examples
-
Vulnerability Exploitation
Weak auth on model hosting or registries enables download.
-
Central Model Registry Risk
A single poorly locked registry becomes a high-value target.
Prevention
-
Access Control
Strong authn/authz, short-lived credentials, and audited downloads.
-
Network Restrictions
Private endpoints, egress controls, and no public weight hosting.
-
Monitoring
Detect unusual inference volume that looks like extraction.
Attack scenarios
-
Model Theft
Stolen weights are reused by competitors or attackers.
-
Insider Leak
Employee access expands the chance of exfiltration.