Browse / AI Framework and Platform Flaws
AI Framework and Platform Flaws
Classic software bugs in the libraries and platforms that run AI: memory corruption, traversal, SSRF, auth. 1,542 records. Reference: CWE.
- Denial of service 377
- Authentication and access control 256
- Path traversal and file access 192
- Memory corruption 159
- Code and command execution 133
- Server-side request forgery 103
- Information exposure 78
- Sandbox escape 76
- Cross-site scripting 68
- Other weaknesses 46
- SQL and query injection 22
- Real-world incidents 10
- Cross-site request forgery 8
- Unsafe deserialisation 8
- Data exfiltration 6
- MLflow Tracking Server Artifact Handler Directory Traversal Remote Code Execution Vulnerability Unauthenticated artifact requests carry unvalidated paths that escape the artifact directory, letting an attacker write files that lead to code execution as the service account.
- Ray dashboard DELETE endpoints allow unauthenticated browser-triggered DoS (Serve shutdown / job deletion) Browser-request middleware only checks POST and PUT, so a DNS-rebound page can send unauthenticated DELETE requests that shut down Serve or delete jobs.
- MindsDB affected by a SSRF vulnerability A filename-cleaning function in the file upload path mishandles input, letting an attacker make the server issue requests to arbitrary destinations.
- Milvus: Unauthenticated Access to Restful API on Metrics Port (9091) Leads to Critical System Compromise A metrics port exposes the full REST API without authentication, and a debug expression endpoint accepts a predictable default token.
- LangChain affected by SSRF via image_url token counting in ChatOpenAI.get_num_tokens_from_messages Token counting for chat messages fetches any image_url in the input, so a crafted URL makes the server issue requests to arbitrary hosts.
- @langchain/community affected by SSRF Bypass in RecursiveUrlLoader via insufficient URL origin validation Crawled pages can link to hosts whose names merely start with the base domain or to private and metadata addresses, which the crawler fetches without validation.
- LangSmith Client SDK Affected by Server-Side Request Forgery via Tracing Header Injection A crafted baggage header makes the LangSmith tracing SDK accept an attacker-chosen endpoint, so run data is sent to that server.
- DJI Romo Cloud Authorization Bug Reportedly Exposed Camera, Microphone, and Home-Mapping Data From Nearly 7,000 Robot Vacuums A DJI robot vacuum cloud authorization flaw let credentials used by a developer's AI-assisted reverse-engineering expose camera, microphone, and map data from nearly 7,000 devices.
- Pydantic AI has Server-Side Request Forgery (SSRF) in URL Download Handling The URL download helper fetches URLs from supplied message history without rejecting internal or metadata addresses.
- Pydantic AI has Stored XSS via Path Traversal in Web UI CDN URL An unvalidated version parameter lets traversal sequences make the server fetch attacker-controlled HTML from the CDN and serve it in the app's origin.
- n8n's domain allowlist bypass enables credential exfiltration Wildcard domain patterns in the credential allowlist fail to restrict domains correctly, so credentials are attached to requests sent to unintended hosts.
- n8n's Unsafe Buffer Allocation Allows In-Process Memory Disclosure in Task Runner Untrusted task runner code can call unsafe buffer allocation and receive uninitialized memory that may hold residual data from the same process.
- n8n Has Expression Escape Vulnerability Leading to RCE Crafted expressions in workflow parameters escape the expression evaluator's restrictions and run operating system commands on the host.
- n8n Vulnerable to Command Injection in Community Package Installation An administrator can inject shell metacharacters into the community package installation flow, so the n8n host runs attacker-chosen commands.
- n8n's Improper CSP Enforcement in Webhook Responses May Allow Stored XSS The CSP sandbox is not applied to some webhook and HTTP responses, so an author's HTML runs with the origin's privileges for other users.
- n8n has a Python sandbox escape An authenticated user can break out of the Python code node sandbox when task runners are enabled, running code outside the intended boundary.
- n8n has OS Command Injection in Git Node Workflow editors can pass input to the n8n Git node that reaches shell commands, letting them run OS commands or read host files.
- n8n's Improper File Access Controls Allow Arbitrary File Read by Authenticated Users Workflow authors can read sensitive files on the host, such as configuration and stored credentials, through file access nodes.
- n8n Merge Node has Arbitrary File Write leading to RCE The Merge node's SQL query mode lets authenticated workflow editors write arbitrary files to the server, which can lead to code execution.
- n8n Vulnerable to Arbitrary File Write on Remote Systems via SSH Node Unauthenticated file uploads routed through workflows to SSH transfers do not validate metadata, letting filenames write to unintended remote paths.
- n8n Has Stored Cross-site Scripting via Markdown Rendering in Workflow UI Markdown rendering in workflow sticky notes lets a stored script run in other users' sessions when they open the workflow.
- Unstructured has Path Traversal via Malicious MSG Attachment that Allows Arbitrary File Write Traversal sequences in MSG attachment filenames make the partitioner write attachments outside the intended output directory.
- H2O has an External Control of File Name or Path vulnerability Attacker-controlled data parsed as a file header is exported through the Frames API to arbitrary server paths, letting the attacker overwrite files such as SSH keys.
- Khoj has an IDOR in Notion OAuth Flow that Enables Index Poisoning The Notion OAuth callback trusts an attacker-supplied state value as the target user ID, binding the attacker's integration to a victim's account.
- llama-index-core vulnerable to Uncontrolled Resource Consumption The directory reader loads every file into memory before applying the file count limit, so large directories exhaust memory.
- Lollms has an Improper Access Control vulnerability Socket.IO generation and cancel events have no authentication, so unauthenticated clients can trigger or interrupt model runs and alter shared server state.
- mlflow Creates of Temporary File in Directory with Insecure Permissions The temporary virtualenv directory is created world-writable, letting a local user race and replace Python files that later run as code.
- SageMaker Python SDK has Insecure TLS Configuration The SDK disabled TLS certificate verification for Triton model downloads, so a network attacker can substitute a malicious model or dependency that then runs in the container.
- vLLM has RCE In Video Processing A crafted video URL sent to a video model makes the server's image decoder overflow the heap, after an info leak reveals memory layout, letting the attacker run commands.
- LobeHub Vulnerable to Improper Authorization in Presigned Upload Client-supplied file name and size fields in the presigned upload request are not validated server-side, allowing writes to unintended paths and misreported storage usage.
- vLLM vulnerable to Server-Side Request Forgery (SSRF) through MediaConnector Host restrictions in the media loader parse backslashes differently in two libraries, so a crafted URL can reach internal hosts.
- n8n Unsafe Workflow Expression Evaluation Allows Remote Code Execution User-authored workflow expressions run in an insufficiently isolated runtime, letting authenticated users execute code as the n8n process.
- BentoML has a Path Traversal via Bentofile Configuration A crafted bentofile can reference paths outside the build context, so building it copies arbitrary host files such as keys into the packaged bento.
- vm2 has a Sandbox Escape A promise callback in the sandbox is sanitised for one Promise realm but not the host-global one, letting sandboxed code reach host process objects.
- Langflow affected by Remote Code Execution via validate_code() exec() Unauthenticated requests to the validation endpoint supply execution globals that the server evaluates, letting attackers run arbitrary code with root privileges.
- Chainlit contain a server-side request forgery (SSRF) vulnerability An authenticated user can supply an element URL that the server fetches, letting it reach internal services and cloud metadata endpoints.
- chainlit : Chainlit allows authenticated arbitrary file reads through custom element updates An authenticated user supplies an arbitrary file path in a custom element update, and the server copies that file into a session for retrieval.
- Lobe Chat has IDOR in Knowledge Base File Removal that Allows Cross User File Deletion A logged-in user can delete knowledge-base files owned by other users because the removal endpoint omits its ownership check.
- Chainlit contains an authorization bypass vulnerability A user-controlled key in Chainlit lets a logged-in user view or take ownership of other users' threads.
- n8n: Webhook Node IP Whitelist Bypass via Partial String Matching The n8n webhook IP whitelist used substring matching, so a client whose address merely contains a trusted entry is accepted.
- Label Studio is vulnerable to full account takeover by chaining Stored XSS + IDOR in User Profile via custom_hotkeys field Stored script injected through a hotkey setting runs in other users' browsers, and missing CSRF protection lets it steal API tokens.
- MindsDB has improper sanitation of filepath that leads to information disclosure and DOS An absolute file path in a JSON upload body escapes the temp directory, so the server reads and moves arbitrary local files.
- MLFlow is vulnerable to DNS rebinding attacks due to a lack of Origin header validation The REST server does not check the Origin header, so a DNS-rebinding page can call endpoints that read, change or delete experiments.
- n8n's Missing Stripe-Signature Verification Allows Unauthenticated Forged Webhooks The Stripe Trigger never checks incoming webhook requests against the stored signing secret, so anyone who knows the URL can start workflows with forged events.
- n8n Vulnerable to Unauthenticated File Access via Improper Webhook Request Handling A publicly reachable form workflow in n8n lets unauthenticated requests read files on the underlying server.
- n8n Vulnerable to RCE via Arbitrary File Write An authenticated user can make n8n write files that lead to untrusted code running on the server.
- Langflow Missing Authentication on Critical API Endpoints Monitoring endpoints for messages and transactions lack the authentication dependency, so anyone can read conversation data or delete messages.
- Feast vulnerable to Deserialization of Untrusted Data Config YAML files are parsed with an unsafe loader that instantiates arbitrary Python objects, so a modified file runs commands.
- n8n's Possible Stored XSS in "Respond to Webhook" Node May Execute Outside iframe Sandbox HTML returned by the Respond to Webhook node escapes the iframe sandbox meant to contain it, so injected scripts run in the editor's top-level window.
- n8n Vulnerable to Arbitrary Command Execution in Pyodide based Python Code Node An authenticated user who can edit workflows writes Python in the Pyodide-based Code node and escapes its sandbox to run commands as the n8n process.