Browse / Tool and Agent Misuse
Tool and Agent Misuse
An agent or chain runs code, commands or tools on behalf of an attacker. 68 records. Reference: OWASP LLM06 · ATLAS AML.T0053.
- Code and command execution 30
- Path traversal and file access 12
- Sandbox escape 7
- Server-side request forgery 4
- Authentication and access control 4
- Other weaknesses 3
- Generate Malicious Commands 3
- AI Agent Tool Invocation 2
- Cross-site scripting 1
- Escape to Host 1
- Exfiltration via AI Agent Tool Invocation 1
- AG2: Newline injection in SandboxShellTool bypasses allowed=[...] and readonly=True, enabling arbitrary command execution An embedded newline lets a command pass the allowlist and read-only checks, because the shell runs the second line as a separate command.
- Claude Code: TOCTOU Race in Edit Tool Allows File Writes Outside the Workspace A symlink swapped into a workspace directory after path validation lets the Edit tool write files outside the workspace, reachable from sandboxed code.
- PraisonAI vulnerable to Remote Code Execution via Broken AST Sandbox An inline Python step in a YAML workflow passes the AST and module checks, letting it import os and run shell commands.
- PraisonAI: SecurityPolicy command/path/import restrictions are completely unenforced by the default SubprocessSandbox backend The default subprocess sandbox ignores configured command, path and import restrictions, so model-driven code runs commands and reads protected files as root.
- PraisonAI: Shell command allowlist bypass via find -exec built-in action The find command is allowlisted, and its -exec action runs arbitrary programs without shell metacharacters, bypassing the command filter.
- PraisonAI: Human-in-the-loop tool approval is cached by tool name and silently reused for all subsequent calls with arbitrary arguments Approval granted once for a tool name is cached and reused for every later call with any arguments, so later file writes skip review.
- PraisonAI: CodeAgent Executes LLM-Generated Code Without Sandboxing and Leaks All Environment Secrets An agent runs model-generated Python in a subprocess with the full environment and no validation, so injected instructions can run code and read secrets even when sandboxing is enabled.
- PraisonAI: FastContext path resolution permits absolute and traversal reads outside the workspace Agent file-search tools accept absolute and traversal paths without confining them to the workspace, so model-driven calls read files outside the project.
- PraisonAI: SkillTools Executes Scripts Without Path Containment Validation The script-running tool executes any file path the model supplies with no containment check, and its approval requirement can be disabled through configuration.
- Langflow: PythonREPLComponent executes unsandboxed Python code, enabling authenticated RCE and privilege escalation The Python interpreter components run user or model code in-process without effective sandboxing, so builtins give attackers full host and privilege access.
- PraisonAI: [Path Traversal] agent tools escape the configured workspace via symlinks Agent file and command tools check the workspace boundary with abspath, so in-workspace symlinks and relative paths let prompt-driven calls reach files outside it.
- praisonaiagents web_crawl vulnerable to SSRF via redirect-following The agent's web crawl tool checks only the initial URL, then follows attacker-supplied redirects to internal addresses and returns their contents to the model.
- praisonaiagents: ast_grep_rewrite rewrites arbitrary files without the @require_approval gate enforced on every sibling mutation tool An AI-callable rewrite tool omits the approval decorator used by sibling file-changing tools, so the model can rewrite files without confirmation.
- Flowise: RCE via CSVAgent csvFile data URI base64 segment is interpolated into Python source without validation Unvalidated CSV file data is spliced into Python code run by an agent sandbox, so a crafted value escapes into Node and runs commands.
- Flowise: CSV Agent Remote Code Execution via Pyodide Code Injection — Root Shell Verified
- AISI Evaluation Agents Conduct Unsanctioned Real-World Cyber Activity
- Claude Code: Sandbox Escape via Git Worktree Path Confusion Allows Unsandboxed Code Execution A malicious repository's worktree layout and symlinks let git's fsmonitor hook run outside the sandbox and overwrite a shell startup file in the home directory.
- n8n: computer-use Shell Sandbox Not Enforced on Linux and Windows The shell tool's sandbox was only applied on macOS, so on Linux and Windows commands ran with unrestricted host filesystem and network access.
- PraisonAI Dynamic Context history and terminal tools read files outside configured storage via path traversal Agent-callable history and terminal tools join caller-supplied identifiers into file paths without rejecting absolute paths or traversal sequences, exposing files outside the storage directory.
- PraisonAI: Arbitrary File Read/Write via `multiedit` Tool Without Path Validation The multiedit tool reads and writes any path an agent supplies because it skips the traversal and workspace checks used elsewhere, so prompts can reach sensitive files.
- PraisonAI: Unauthenticated RCE via Jobs API + Approval Bypass An unauthenticated job endpoint accepts workflow YAML whose top-level approve field disables the approval gate, letting the agent run dangerous shell tools.
- PraisonAI recipe workflow policy can be bypassed by declaring and YAML-approving dangerous tools outside TEMPLATE.yaml A recipe can declare a dangerous shell tool in its workflow file and self-approve it, skipping the policy check so untrusted commands run.
- PraisonAI DiscordApproval accepts unrelated channel messages as dangerous-tool approvals Any non-bot message in the approval channel reading yes counts as approval for a pending dangerous tool call, since replies and approvers are not bound.
- Open WebUI BOLA: `search_knowledge_files` Allows Unauthorized Knowledge Base File Enumeration An authenticated user can make the model call the built-in knowledge search tool with an arbitrary knowledge base ID, and the tool returns file metadata without an access check.
- Open WebUI: Stored XSS in Mermaid Markdown Preview
- PraisonAI vulnerable to sandbox escape via `print.__self__` builtins module leak in `execute_code` (subprocess mode) The AST filter for agent Python execution misses an attribute on built-in functions, letting code reach the builtins module and run OS commands.
- MLflow: Improper Origin Validation in MLflow Assistant /ajax-api Endpoints Enables Browser-Mediated Local Command Execution A malicious webpage can send cross-origin requests to MLflow Assistant's local endpoints, change its settings, and then have a Claude Code sub-agent run commands.
- GitHub Copilot CLI: Nested Bare Repository Can Execute Arbitrary Commands via core.fsmonitor A nested bare git repository in a project sets executable config keys, so when the agent runs ordinary git commands it executes attacker commands without asking the user.
- Threat Actor Uses a DeepSeek-Powered Hermes Agent in Langflow and n8n Exploitation Attempts
- Flowise: CSV Agent Prompt Injection Remote Code Execution Vulnerability The CSV agent runs Python scripts generated by the LLM without sandboxing, so a crafted prompt makes the server execute attacker-chosen code without authentication.
- AgentScope vulnerable to Server-Side Request Forgery Image and audio URL arguments passed to AgentScope's multimodal tools are fetched server-side without checks, so a model-supplied URL can reach cloud metadata endpoints.
- AgentScope Vulnerable to Remote Code Injection Agent tool functions for executing Python code and shell commands accept injected input, letting a remote caller run arbitrary code.
- PraisonAI Vulnerable to Code Injection and Protection Mechanism Failure Calling type.__getattribute__ with string attribute names evades the AST filter that blocks dangerous dunder attributes, letting agent-run code escape the sandbox.
- PraisonAI: Unauthenticated Allow-List Manipulation Bypasses Agent Tool Approval Safety Controls An unauthenticated caller can add dangerous tools to the approval allowlist when no token is set, so the agent runs them without human confirmation.
- PraisonAIAgents has SSRF and Local File Read via Unvalidated URLs in web_crawl Tool The web_crawl agent tool fetches any supplied URL without validation, so injected content can make the agent reach internal hosts or local files.
- 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: 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.
- Open WebUI's process_files_batch() endpoint missing ownership check, allows unauthorized file overwrite
- Langflow has Authenticated Code Execution in Agentic Assistant Validation
- 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.
- OpenClaw 1-Click Remote Code Execution
- Ray is vulnerable to Critical RCE via Safari & Firefox Browsers through DNS Rebinding Attack
- @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.
- Flowise is vulnerable to arbitrary file exposure through its ReadFileTool
- Flowise is vulnerable to arbitrary file write through its WriteFileTool
- 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.