Browse / Indirect Prompt Injection
Indirect Prompt Injection
Instructions hidden in documents, web pages, emails or tool output that the model later reads. 20 records. Reference: OWASP LLM01 · ATLAS AML.T0051.
- Code and command execution 5
- Prompt Infiltration via Public-Facing Application 4
- Server-side request forgery 3
- Data exfiltration 2
- LLM Prompt Injection: Indirect 2
- Other weaknesses 2
- Path traversal and file access 1
- AI Agent Clickbait 1
- praisonaiagents vulnerable to SSRF in web_crawl tool via redirect-following and DNS rebinding (validate-then-fetch gap)
- PraisonAI: Server-Side Request Forgery (SSRF) in SearxNG / search_web tools via attacker-controlled searxng_url parameter Instructions in fetched web or file content can steer the agent's search tool to attacker-chosen internal URLs, whose responses are returned.
- MCP Server Kubernetes: kubectl-generic flag injection enables Kubernetes bearer token exfiltration Injected instructions in pod logs make the agent pass attacker server and insecure flags to kubectl, sending the operator's bearer token elsewhere.
- PraisonAIAgents: SSRF via unvalidated URL in `web_crawl` httpx fallback
- PraisonAI: Shell Injection in run_python() via Unescaped $() Substitution
- Claude SDK for TypeScript: Memory Tool Path Validation Allows Sandbox Escape to Sibling Directories A prompt-injected model supplies a path that shares the memory root's name as a string prefix, so the prefix check lets reads and writes reach a sibling directory.
- GitHub Copilot CLI Dangerous Shell Expansion Patterns Enable Arbitrary Code Execution
- Claude Code Vulnerable to Command Injection via Directory Change Bypasses Write Protection Untrusted content in context could make Claude Code use a cd command to enter protected directories and write files without the user being asked to confirm.
- ZombieAgent: Data Exfiltration Attack on ChatGPT
- @akoskm/create-mcp-server-stdio is vulnerable to MCP Server Command Injection through `exec` API Prompt-injected input to an MCP tool is interpolated into a shell exec call, so shell metacharacters run extra commands on the server.
- mcp-markdownify-server vulnerable to command injection in pptx-to-markdown tool Instructions hidden in a Markdown file make the MCP client call the pptx tool with shell metacharacters that the server passes to child_process.exec.
- AI ClickFix: Hijacking Computer-Use Agents Using ClickFix
- Data Destruction via Indirect Prompt Injection Targeting Claude Computer-Use
- Data Exfiltration from Slack AI via Indirect Prompt Injection
- Planting Instructions for Delayed Automatic AI Agent Tool Invocation
- Google Bard Conversation Exfiltration
- Indirect Prompt Injection Threats: Bing Chat Data Pirate
- Prompt Injection via GitHub Patch in Brave AI Chat (Leo)
- Source Code and data exfiltration via Github Copilot
- LLM01: Invisible Prompt Injection