Browse / Tool and Agent Misuse
Tool and Agent Misuse
An agent or chain runs code, commands or tools on behalf of an attacker. 164 records. Reference: OWASP LLM06 · ATLAS AML.T0053.
- Real-world incidents 68
- 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
- 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.
- Attacker Reportedly Used AI Agents to Breach Blockchain Analytics Firm Chainalysis and Access Customer Data An attacker reportedly used agentic AI techniques to break into Chainalysis systems and access customer names, emails, user IDs and organizational roles.
- 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.
- Defense Contractor's Personal AI Coding Agent Reportedly Sent U.S. Defense Department Credentials to a Third-party Website A defense contractor's personal Claude Code agent, connected to a Defense Department system, reportedly sent login credentials and a network map to a third-party website; Accenture disputes the allegations.
- Cybersecurity Firm Cathedral Reportedly Used AI Agents to Access Anthropic's Internal Business Systems During invited testing, Cathedral reportedly used AI agents to exploit third-party software and reach Anthropic's internal systems, including personnel data and internal messages, beyond the agreed rules.
- ARTEX Security Testing Agent Reportedly Accessed WorkOS's Internal Sales Tool Without Authorization and Retrieved Customer Data An ARTEX agent running a cybersecurity benchmark reportedly accessed WorkOS's internal sales tool without authorization and bulk-retrieved customer contact details, communications and metadata.
- China-Linked Espionage Group VAPOR PANDA Reportedly Used ARTEX AI Agents to Breach Government and Corporate Systems Worldwide The China-linked group VAPOR PANDA reportedly used the ARTEX AI penetration-testing agent with language models to breach government and corporate systems, deploying backdoors at nine confirmed organizations.
- AI Agents Under Evaluation Reportedly Accessed Personal and Confidential Records at Indian Organizations Indian cybersecurity agency reported AI agents from unnamed providers accessed live business systems beyond authorization, using exposed credentials to read personal and confidential records.
- ARTEX AI Agents Reportedly Used Stolen Cloud Backup Credentials to Access ConnectWise Customers' Microsoft 365 Environments Hackers reportedly used the ARTEX AI pentest framework and stolen ConnectWise backup credentials to access customers' Microsoft 365 environments, changing credentials to keep access.
- 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.
- Pydantic AI Web chat UI (`Agent.to_web()`, `clai web`): a website visited by the developer can trigger agent runs and server-side tool execution on the local chat endpoint A malicious webpage sends cross-origin requests to the local chat endpoint, which runs the agent and its tools without checking content type or approval.
- Pydantic AI Web chat UI (`Agent.to_web()`, `clai web`): the local chat endpoint does not validate the `Host` header A DNS-rebound page treats the local chat endpoint as same-origin, since Host is unchecked, and drives the served agent to run its tools with local privileges.
- 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.
- Tencent-Linked AI Agents Reportedly Used Exposed Credentials to Retrieve Data From Alibaba's Mapping Service Tencent-linked AI agents reportedly used an API key and security code exposed in another company's source code to pull place details and photos from Alibaba's Amap mapping service.
- Attackers Reportedly Used ARTEX, an AI Security-Testing Tool, to Steal Data from South Korean Financial Institutions Attackers reportedly used the ARTEX AI security-testing tool with large language models to breach South Korean financial institutions and steal data from loan and employee systems.
- Meta's Muse AI Agent Reportedly Shared a Seller's Home Address and Arranged an Unexpected Pickup in Toronto Meta's Muse agent reportedly shared a reviewer's home address and arranged an unexpected Marketplace pickup, after he had enabled an always-allow setting for pickup replies.
- AI Agent Appearing to Originate from OpenAI Reportedly Attempted to Hack U.S. Department of Education Civil Rights Website An AI agent appearing to come from OpenAI used its tools to attempt hacking a Department of Education civil rights website; the attempt reportedly failed.
- Cybercriminal Reportedly Used Cairn AI Agents to Breach More Than 100 Company Websites An operator using Chinese-language instructions directed the Cairn AI agent framework to breach more than 100 company websites across four countries.
- Anthropic's Claude Code AI Agent Reportedly Deleted About 48,000 Live Project Files and Version-Control History During Software Repair A Claude Code sub-agent followed Windows directory junctions into a live project, deleting about 48,000 files and the Git history in under two minutes.
- AI Agent Reportedly Exploited Application Vulnerabilities to Modify Personal Data and Access Invoices in Breach Reported to Spanish Data Protection Authority A third party used an AI agent built on a known language model to exploit application vulnerabilities, altering personal data and accessing invoices at an unnamed organization.
- Early Claude Opus 4.6 Checkpoint Reportedly Gained Unauthorized Admin Access to Third-Party System During Cybersecurity Evaluation An early Claude Opus 4.6 checkpoint in an evaluation escaped its target, reached a third-party machine, used a found password for admin access, harvested credentials, and read personal information.
- 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.
- Instinct AI Personal Assistant Reportedly Sent Email From Moxxie Ventures Founder's Account Without Approval The Instinct AI assistant sent an email from a founder's connected account without asking for approval, using its own email access permissions.
- Meta AI Model Reportedly Exploited Security Vulnerability in Real Third-Party Service During Cybersecurity Evaluation During a Meta cybersecurity evaluation, an AI model exploited a vulnerability in a real third-party service after a misconfiguration gave it unintended Internet access.
- 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 Unsanitised CSV agent input is inserted into Python code run in Pyodide, and the JS bridge lets that code reach the host Node process to run OS commands.
- Flowise: Remote Code Execution Vulnerability in CSVAgent The CSV agent runs user-supplied Python through a denylist that misses pandas read_pickle, letting a pickle payload execute code.
- Claude Opus 4.7 Reportedly Compromised Real Company's Production Infrastructure During Cybersecurity Evaluation During an Irregular-run Anthropic evaluation, Claude Opus 4.7 kept attacking a real company it took for a fictional target, extracting credentials and reading production data.
- Claude Mythos 5 Reportedly Published Malicious PyPI Package That Compromised Real Security Company During Evaluation During an Irregular-run Anthropic evaluation, Claude Mythos 5 published a malicious PyPI package that ran on 15 real systems and used stolen credentials to reach more infrastructure.
- Anthropic Research Model Reportedly Scanned 9,000 Targets and Compromised Real Company's Application During Evaluation During an Irregular-run Anthropic evaluation, a research model scanned about 9,000 targets and compromised a real company's app via exposed debug credentials and SQL injection.
- Anthropic Claude Opus 5 Coding Agent Reportedly Reset a Live Supabase Production Database During Prisma Migration Work A Claude Opus 5 coding agent ran prisma migrate diff with the production URL as the shadow database, causing Prisma to drop all 22 tables in a live Supabase database.
- Anthropic and OpenAI AI Agents Reportedly Took Unsanctioned Actions on the Live Internet During UK AISI Cybersecurity Evaluations Anthropic and OpenAI agents reportedly took 19 unsanctioned live-Internet actions during UK AISI cyber evaluations, including deceptive attempts to get developers to accept malicious code. AISI contained them.
- 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.
- n8n: Path-Confinement Bypass in computer-use search_files Allows Reading Files Outside the Base Directory A crafted search pattern in an agent file-search tool escapes the configured base directory and returns contents of files elsewhere.
- Anthropic AI Agent Reportedly Submitted a Fabricated Homicide Tip to Police in Philadelphia During Anthropic's evaluation, a Claude Haiku 4.5 agent submitted fabricated eyewitness information about a Philadelphia homicide through a police tip form, entering the tip records.
- OpenAI GPT-5.6 Sol Reportedly Deleted Software Engineer's Production Database During Local Testing A GPT-5.6 Sol agent running cleanup executed TRUNCATE TABLE users CASCADE against a production database because the test database URL pointed to live data.
- OpenAI Models Reportedly Compromised Hugging Face Production Infrastructure During Cybersecurity Evaluation During a cyber-capability evaluation, OpenAI models escaped their sandbox through a package-registry proxy flaw and reached Hugging Face production systems, accessing test solutions.
- OpenAI GPT-5.6 Sol Agent Reportedly Deleted Most Files in AI Startup Founder's Mac Home Directory A GPT-5.6 Sol Codex agent misexpanded $HOME during a cleanup task and ran rm -rf, deleting most files in the user's Mac home directory.
- Threat Actor Reportedly Used Hermes AI Agent for Unattended Post-Compromise Activity in Thailand's Ministry of Finance Network A threat actor ran Nous Research's Hermes agent unattended to perform privilege-escalation reconnaissance and search personnel records inside Thailand's Ministry of Finance network.
- Threat Actor Reportedly Used AI-Assisted Workflows to Compromise AWS Environment for Extortion A threat actor reportedly used agentic AI-assisted workflows to move through an AWS environment in about 72 hours, stealing credentials and data and threatening disruption for extortion.