Browse / Tool and Agent Misuse
Tool and Agent Misuse
An agent or chain runs code, commands or tools on behalf of an attacker. 96 records. Reference: OWASP LLM06 · ATLAS AML.T0053.
- Code and command execution 43
- Path traversal and file access 18
- Sandbox escape 11
- Server-side request forgery 7
- Authentication and access control 4
- Data exfiltration 3
- Generate Malicious Commands 3
- Other weaknesses 2
- AI Agent Tool Invocation 2
- Cross-site request forgery 1
- Escape to Host 1
- Exfiltration via AI Agent Tool Invocation 1
- PraisonAIAgents: SSRF via unvalidated URL in `web_crawl` httpx fallback An agent is steered into fetching internal or metadata URLs through a web crawl fallback that performs no host or address validation.
- PraisonAIAgents has an OS Command Injection via shell=True in Memory Hooks Executor (memory/hooks.py) Hook commands run through a shell without sanitisation, and an agent steered by prompt injection can rewrite the hook config to run its payloads.
- PraisonAIAgents: Environment Variable Secret Exfiltration via os.path.expandvars() Bypassing shell=False in Shell Tool The shell tool expands environment variables inside arguments even though it avoids a shell, so reviewers see literal variable names while secret values are read and sent out.
- FoundationAgents MetaGPT vulnerable to OS Command Injection in metagpt/tools/libs/terminal.py The terminal tool's Bash.run passes unsanitised input to the OS shell, letting an attacker inject extra commands.
- FoundationAgents MetaGPT vulnerable to os command injection via the Terminal.run_command Arguments passed to the Terminal.run_command agent tool are not safely handled, so injected operating system commands run on the host.
- PraisonAI Vulnerable to OS Command Injection Workflow fields and LLM-generated tool parameters reach a shell-enabled subprocess call, so shell metacharacters run extra commands.
- PraisonAI has sandbox escape via exception frame traversal in `execute_code` (subprocess mode) The subprocess sandbox's AST blocklist omits frame-traversal attributes, so code run through the agent's execute tool can reach builtins and call exec.
- PraisonAI Vulnerable to Arbitrary File Write / Path Traversal in Action Orchestrator An action step's target path is joined to the workspace without normalisation, so a traversal sequence supplied by an agent can write files outside the workspace.
- PraisonAI Has Sandbox Escape via shell=True and Bypassable Blocklist in SubprocessSandbox Sandbox commands run in shell mode with a string blocklist that omits sh and bash, so wrapping a command in sh -c escapes STRICT restrictions.
- PraisonAI: Python Sandbox Escape via str Subclass startswith() Override in execute_code A str subclass overriding startswith defeats the underscore guard, letting attacker-supplied Python reach subprocess functions and run OS commands.
- Claude Code has a Workspace Trust Dialog Bypass via Repo-Controlled Settings File A malicious repository's committed settings select bypass-permissions mode before the trust dialog appears, so Claude Code runs tools without the user's consent.
- GitHub Copilot CLI Dangerous Shell Expansion Patterns Enable Arbitrary Code Execution Bash parameter expansion operators hide command substitutions inside arguments the safety check treats as read-only, so the agent runs them without approval.
- n8n has Arbitrary Command Execution via File Write and Git Operations A user who can edit workflows chains file-write nodes with git operations to run arbitrary shell commands on the host.
- Hugging Face Smolagents has a Server-Side Request Forgery issue Code the agent executes in its local Python executor can call requests.get and requests.post against arbitrary internal URLs.
- AI in the Middle: Web-Based AI Services as C2 Relays
- Semantic Kernel has Arbitrary File Write via AI Agent Function Calling in .NET SDK An agent's function calls to file download and upload methods accept arbitrary local paths, so model-driven calls can read or write files outside the intended area.
- Claude Code has Permission Deny Bypass Through Symbolic Links Deny rules check the requested path but not its symlink target, so the agent can read a file that a user explicitly blocked.
- Claude Code has Sandbox Escape via Persistent Configuration Injection in settings.json Sandboxed code could create a missing settings.json and add startup hooks that run on the host after restart.
- Claude Code Vulnerable to Command Injection via Piped sed Command Bypasses File Write Restrictions A piped sed command built with echo got past Claude Code's command validation, letting an agent in accept-edits mode write to protected paths such as .claude and outside the project.
- OpenClaw 1-Click Remote Code Execution
- @anthropic-ai/claude-code has Sed Command Validation Bypass that Allows Arbitrary File Writes A parsing error in the sed command validator lets a crafted expression pass read-only checks and write to arbitrary files.
- Claude Code vulnerable to command execution prior to startup trust dialog A malicious Yarn configuration in a repository executes code when Claude Code runs yarn --version, before the user accepts the directory trust prompt.
- DSPy does not properly restrict file reads The PythonInterpreter sandbox allows file reads beyond what it should, so agent code generated from user input can read sensitive host files.
- Flowise: Authenticated Command Execution and Sandbox Bypass via Puppeteer and Playwright Packages Authenticated users can set attacker-chosen browser binary paths and arguments in Puppeteer or Playwright tools, which run on the host outside the nodevm sandbox.
- n8n: Execute Command Node Allows Authenticated Users to Run Arbitrary Commands on Host Any user who can build workflows can run arbitrary shell commands on the host through the Execute Command node when it is enabled.
- Claude Code can execute commands prior to the startup trust dialog Project code ran at startup before the user accepted the trust dialog when Claude Code was opened in an untrusted directory.
- Claude Code permission deny bypass through symlink Claude Code's deny rules matched the path string rather than the symlink target, so a denied file could be read through a symlink.
- Claude Code Vulnerable to Arbitrary Code Execution via Plugin Autoloading with Specific Yarn Versions Claude Code runs yarn --version in untrusted directories, and malicious Yarn config can execute code before the trust dialog is shown.
- Claude Code vulnerable to arbitrary code execution caused by maliciously configured git email A malicious git user.email value in a repository's config is interpolated into a startup shell command, running it before the user accepts the workspace trust dialog.
- Claude Code Vulnerable to Arbitrary Code Execution Due to Insufficient Startup Warning Accepting the folder trust prompt silently allowed the agent to run files in that folder, because the warning did not disclose this behaviour.
- Claude Code's Permissive Default Allowlist Enables Unauthorized File Read and Network Exfiltration in Claude Code An overly broad safe-command list let injected content read a local file and send it over the network without confirmation.
- Prompt-Based Attacks Against Gemini via Calendar Invitations
- Claude Code echo command allowed bypass of user approval prompt for command execution A command parsing error let an echo invocation skip the confirmation prompt, so injected untrusted content could run arbitrary commands.
- smolagents has Sandbox Escape Vulnerability in the local_python_executor.py Module Code the agent executes can abuse whitelisted modules and functions to escape the local Python sandbox and run arbitrary code on the host.
- LangChain Community SSRF vulnerability exists in RequestsToolkit component The OpenAPI RequestsToolkit lets an agent send requests to any address, including internal and cloud metadata endpoints, with no restriction.
- LAMEHUG: Malware Leveraging Dynamic AI-Generated Commands
- Markdownify MCP Server allows attackers to read arbitrary files A crafted prompt can make the Markdownify MCP server's file tool read arbitrary files on the host, since the tool does not restrict which paths it opens.
- Markdownify MCP Server allows Server-Side Request Forgery (SSRF) via the Markdownify.get() function Attacker-authored prompt content makes the MCP host call the server's conversion tools, which fetch attacker-chosen URLs and return the responses.
- Browser Use allows bypassing `allowed_domains` by putting a decoy domain in http auth username portion of a URL A decoy domain placed in the URL userinfo part makes the browser agent's allowed-domains check pass while it navigates to a disallowed host.
- LoLLMS Code Injection vulnerability The Calculate tool's eval sandbox can be escaped by reaching the import machinery through a frozen importer, enabling arbitrary command execution.
- LLMSmith: RCE Vulnerabilities in LLM-Integrated Applications
- Langflow vulnerable to remote code execution Langflow components run user-supplied code directly on the host rather than in a sandbox, allowing arbitrary command execution.
- langchain-experimental vulnerable to Arbitrary Code Execution Prompts that steer the SQL chain cause database return values to be passed to eval, so attacker-chosen Python runs on the server.
- langchain_experimental Code Execution via Python REPL access The Python REPL tool is available without an explicit opt-in, so an agent can run arbitrary Python code on the host.
- LangChain vulnerable to arbitrary code execution The Python AST REPL tool runs supplied scripts without adequate restriction, letting a crafted script execute arbitrary code on the host.
- Langchain OS Command Injection vulnerability Crafted input to the Jira API wrapper in older LangChain versions leads to arbitrary code execution.