Browse / Improper Output Handling
Improper Output Handling
Model output trusted by downstream code: XSS, SQL injection, SSRF or code execution through what the model wrote. 36 records. Reference: OWASP LLM05.
- Cross-site scripting 15
- Code and command execution 12
- SQL and query injection 6
- Path traversal and file access 1
- Data exfiltration 1
- Denial of service 1
- Langflow: Prompt injection in Langflow Smart Transform can lead to code execution Smart Transform eval()s a lambda the LLM returns after weak format checks, so injected instructions can make the model produce executable Python.
- Flowise: Pyodide validator Unicode homoglyph bypass leads to RCE A regex blacklist on generated Python misses Unicode homoglyph identifiers that Python normalises to dunder names, so forbidden access to builtins runs in the sandbox.
- Open WebUI: Stored web worker XSS via Pyodide A chat message carrying Python code runs in a same-origin worker when a victim clicks Run, letting it make authenticated requests as that user.
- Open WebUI vulnerable to stored XSS via unescaped markdown token in MarkdownTokens.svelte leading to full account takeover and RCE via functions Chat messages containing an iframe-like tag pass a loose string check and are rendered as raw HTML, running attacker script in viewers' sessions.
- PraisonAI: HTTPApproval dashboard renders tool arguments as raw HTML, allowing approval-page XSS to approve dangerous tools Attacker-influenced tool arguments are inserted unescaped into the HTTP approval dashboard, so a script there can approve a dangerous pending tool call.
- PraisonAI: IMAP Command Injection via Unsanitized Email Search Parameters Model-supplied search arguments are concatenated into IMAP SEARCH strings, so quote characters let an attacker add commands that read or alter other mail folders.
- Aider is vulnerable to Code Injection via editor_coder.run function Model-generated content from Architect mode reaches an editor code path without sanitisation, so attacker-influenced text is executed as code.
- Open WebUI has stored XSS via the HTML renedering view Model-generated HTML is rendered in an iframe whose sandbox permits scripts and same-origin access, so injected script can read the session token.
- Open WebUI has stored XSS in Excel file preview
- Flowise: Cypher Injection in GraphCypherQAChain User text flows into an LLM-generated Cypher query that runs unsanitised against Neo4j, letting an attacker run arbitrary graph queries.
- PraisonAI Vulnerable to Stored XSS via Unsanitized Agent Output in HTML Rendering (nh3 Not a Required Dependency) Agent output is converted to HTML and embedded unescaped, and the optional sanitizer returns it unchanged when missing, so injected scripts run for viewers.
- FoundationAgents MetaGPT vulnerable to eval injection Model-generated XML content is passed into eval during ActionNode.xml_fill, so crafted model output causes arbitrary code to run on the server.
- Hugging Face Smolagents has an Injection issue The local Python executor that runs agent-generated code fails to block some constructs, so crafted code escapes its restrictions and runs commands.
- New API has Potential XSS in its MarkdownRenderer component Model-generated HTML is rendered without sanitisation in the chat UI, so a reply containing script runs in the browser and persists in stored history.
- Lobe Chat affected by Cross-Site Scripting(XSS) that can escalate to Remote Code Execution(RCE) Model-generated Mermaid artifact labels containing HTML are rendered unsanitised, so injected script runs in the app and can reach a desktop MCP endpoint.
- Malicious website can execute commands on the local system through XSS in the OpenCode web UI Unsanitised markdown from model responses runs script on the local web UI origin, which then calls a process-spawning API to run commands.
- Open WebUI Affected by an External Model Server (Direct Connections) Code Injection via SSE Events A malicious external model server sends SSE execute events that the frontend runs as JavaScript, so server-supplied content executes in users' browsers.
- Hugging Face Smolagents XPath injection vulnerability in the search_item_ctrl_f function The search helper concatenates its input into an XPath expression, so crafted search text passed by the agent alters the query and reaches unintended page elements.
- Flowise vulnerable to XSS Unsanitised chat input and external web content returned by a custom function are rendered as HTML, so embedded scripts run in other users' browsers.
- Flowise Stored XSS vulnerability through logs in chatbot Chat prompts containing unsanitised HTML are stored in logs and rendered for administrators, so embedded script runs in the admin's browser.
- vLLM has remote code execution vulnerability in the tool call parser for Qwen3-Coder The Qwen3-Coder tool-call parser calls eval() on tool parameters of unrecognised type, so model-generated arguments containing Python code run on the server.
- Open WebUI Vulnerable to a Session Fixation Attack Markdown images in chat content make an admin's browser request an attacker URL, and session cookies lacking Secure and SameSite protections are sent with it.
- @langchain/community SQL Injection vulnerability A crafted user question steers the LLM into generating malicious Cypher, which the chain executes without validation, enabling data modification and theft.
- Open WebUI Stored Cross-Site Scripting Vulnerability A crafted prompt makes the model emit JavaScript that the web interface renders unsanitised in the victim's page.
- Local file inclusion in gradio Output parsed as JSON containing a path key makes the server copy that local file into the cache, where it can later be fetched.
- SQL injection in llama-index Natural-language input is turned into SQL by the text-to-SQL engines and executed without restriction, so prompts can produce destructive queries.
- Langchain vulnerable to arbitrary code execution via the evaluate function in the numexpr library Attacker-influenced text reaching numexpr's evaluate function in LangChain runs arbitrary code, because the library passes that text to an unsafe evaluator.
- LangChain vulnerable to arbitrary code execution Crafted input to helper chains makes the model generate Python that the chain runs without adequate checks, letting a remote party execute arbitrary code.
- llama-index vulnerable to arbitrary code execution Query-engine code generated by the model is passed to Python exec, so crafted model output or prompts can run arbitrary code on the host.
- langchain vulnerable to arbitrary code execution PALChain passes model-generated Python to exec, so crafted prompts can make the host run arbitrary code.
- langchain SQL Injection vulnerability Attacker-influenced input reaching the SQL chain produces injected queries that run against the database and return sensitive records.
- langchain arbitrary code execution vulnerability Python code generated by the LLM is passed to exec in PALChain, so model output runs as code on the host.
- Reflected XSS in AI Chat Bot Greetings at help.shopify.com via Markdown Image Rendering
- AI Playground XSS to steal user-chat messages and access to connected MCP Server
- Null Pointer Dereference by Crafted Response from AI Model
- Reflected XSS in chatbot