CVE Database

SadTalker contains an OS command injection vulnerability in the video muxing process where uploaded audio filenames are interpolated into ffmpeg commands without proper escaping. Attackers can upload audio files with shell metacharacters in the filename to break out of quoted arguments and execute arbitrary system commands when video generation occurs.

FastChat contains an authentication bypass vulnerability in the /register_worker endpoint that allows unauthenticated attackers to register arbitrary worker addresses and perform server-side request forgery. Attackers can register malicious workers under victim model names to intercept user prompts, images, and responses, or probe internal network ports across the worker mesh.

LaVague 0.2.35 contains a remote code execution vulnerability in PythonFromMarkdownExtractor.extract_as_object that evaluates untrusted language model output derived from web page content. Attackers can inject malicious Python code through web pages using indirect prompt injection to execute arbitrary code on the operator's host without review.

Chatbot UI contains an authorization bypass vulnerability in the retrieval endpoint that allows authenticated attackers to access private file content belonging to other users by supplying arbitrary file UUIDs. The endpoint uses a service-role Supabase client that bypasses row-level security and fails to validate file ownership, enabling attackers to retrieve indexed content chunks from victim files through crafted POST requests.

Nightingale (n9e), as of commit 8362cbe (main branch, confirmed 2026-08-27), contains a server-side request forgery vulnerability in the isPublicIP function in aiagent/tools/http.go, the SSRF guard for the http_fetch AI-agent tool. The function only unwraps standard IPv4-mapped (::ffff:a.b.c.d) IPv6 addresses before checking them against the forbidden-range list, and does not classify 6to4 (2002::/16), NAT64 (64:ff9b::/96, 64:ff9b:1::/48), or deprecated site-local (fec0::/10) addresses. On a dual-stack or NAT64-enabled host, an attacker able to supply a URL to the http_fetch tool can bypass the guard by encoding a forbidden IPv4 address (such as the cloud instance-metadata endpoint 169.254.169.254) in one of these IPv6 forms to reach internal or metadata services.

MegaParse 0.0.55 contains an unauthenticated server-side request forgery vulnerability in the POST /v1/url endpoint that fetches caller-supplied URLs server-side. Attackers can supply internal service URLs or metadata endpoints without authentication to read their responses directly from the JSON response.

Plandex 2.2.1 contains a path traversal vulnerability in the ApplyFiles function that allows attackers to write files outside the project directory. Attackers can influence model output through poisoned repository files or attacker-controlled context to write to arbitrary locations like shell rc or cron files, achieving code execution.

llmware 0.4.6 contains an SQL injection vulnerability in the collection-database layer (llmware/resources.py) where filter and lookup values are directly string-interpolated into SQL WHERE clauses without parameterization or escaping, in both the SQLite and PostgreSQL backends. The filter validator only checks keys against an allow-list and never sanitizes values. Attacker-controlled filter values reaching the public API via Library.block_lookup and Query.text_query_with_custom_filter / text_query_by_author_or_speaker can neutralize the intended filter to disclose rows the caller was scoped out of (cross-document/cross-collection disclosure); on PostgreSQL the flaw permits boolean- and UNION-based SQL injection.

TEN Framework 0.11.71 contains unauthenticated arbitrary file read and write vulnerabilities in the TMAN Designer file-content API endpoints. Attackers can submit POST and PUT requests to the /api/designer/v1/file-content endpoints to read arbitrary files or write malicious content to system paths, enabling code execution through authorized_keys, cron files, or executable graph files.

surya 0.22.1 screenshot server contains an unauthenticated arbitrary file read vulnerability in the /info, /page, and /process routes that accept raw file_path parameters. Attackers can read any image or PDF file on the host by supplying arbitrary file paths to Image.open or pypdfium2.PdfDocument, obtaining rendered contents as base64 and using /info as an existence oracle.

ms-swift 4.5.2 contains a server-side request forgery vulnerability in the swift deploy OpenAI-compatible API that fetches multimodal media URLs without validation or redirect filtering. Unauthenticated attackers can supply arbitrary image_url, audio_url, or video_url parameters to make the server issue requests to internal services and cloud metadata endpoints.

AgentScope through 2.0.7.post1 contains a path traversal vulnerability in LocalWorkspace.add_skill that copies arbitrary server directories into the agent workspace via an unconfined source path parameter. Attackers can supply any directory path in the skill_path request parameter to copy files into the skills directory, making them accessible through the workspace skill listing.

marker through 2.0.0 contains a path traversal vulnerability in the FastAPI /marker/upload handler that fails to sanitize the file.filename parameter. Unauthenticated attackers can supply filenames containing directory traversal sequences to write arbitrary files to any location or delete existing files on the system.

Dub contains an open redirect vulnerability in the redir_url query parameter that is accepted on every short link without validation or domain allowlist enforcement. Attackers can append the redir_url parameter to any short link to redirect visitors to arbitrary external URLs through the trusted Dub domain, bypassing destination blacklists and potentially enabling phishing attacks with link cloaking enabled.

OWL's DocumentProcessingToolkit contains a server-side request forgery vulnerability in the extract_document_content tool that fetches caller-supplied URLs with no scheme, host, or IP filtering. Attackers can inject malicious URLs through prompt injection to make the server fetch internal resources, with responses returned to the agent context.

aider (aider-chat) automatically loads a .aider.conf.yml configuration file from the root of the git repository it is launched in. A crafted repository can set test-cmd (executed at startup) or lint-cmd (executed on the first file edit), which aider runs through a shell (subprocess with shell=True) without any user confirmation, LLM interaction, or API key. Consequently, a user who clones and runs aider inside an attacker-supplied repository achieves arbitrary command execution on their machine. The behavior is long-standing and was confirmed on 0.86.3.dev (current main).

LLaMA-Factory contains a server-side request forgery vulnerability in the OpenAI-compatible API multimodal media URL handler that allows unauthenticated attackers to bypass SSRF validation. The check_ssrf_url guard validates URLs once but requests.get follows redirects and re-resolves DNS without re-validation, enabling attackers to use HTTP redirects or DNS rebinding to access internal addresses and cloud metadata endpoints.

zerox 1.1.20 contains an OS command injection vulnerability in the file download mechanism where the temporary file extension derived from document URLs is interpolated unsanitized into shell commands executed by poppler utilities. Attackers can craft document URLs with malicious file extensions containing command substitution syntax to execute arbitrary OS commands before document processing occurs.

QAnything 2.0.0 contains an authentication bypass vulnerability in the /api/local_doc_qa/get_file_base64 and /api/local_doc_qa/get_doc endpoints that allows unauthenticated attackers to access any uploaded file or document. Attackers can enumerate file identifiers through unauthenticated endpoints and retrieve base64-encoded files or parsed document chunks without ownership verification to disclose cross-tenant knowledge base content.

tokenizers (Hugging Face) is affected by an out-of-bounds buffer access in BpeBuilder::build (tokenizers/src/models/bpe/model.rs). When loading a tokenizer.json via Tokenizer::from_file/from_str, the builder sizes a scratch buffer to the longest vocabulary key, then writes each concatenated merge rule into it. A merge whose concatenated token exceeds the longest vocabulary key overruns the buffer, which Rust turns into a panic that aborts the process in Rust and FFI embeddings. This occurs at load time with no encoding required, so an attacker who supplies a crafted tokenizer.json can cause a denial of service. A secondary defect at the same location can cause a usize underflow (panic in debug, potential memory corruption in release) when continuing_subword_prefix is set and a merge token is shorter than the prefix. Observed in version 0.23.1.

potpie through 2.0.0 fails to verify user ownership on the POST /conversations/{conversation_id}/code-changes/sync endpoint. Authenticated attackers can write arbitrary file changes into other users' conversations by supplying their conversation IDs, allowing unauthorized modification of pending changes.

Xinference (affected commit 4a94832, v3.x) contains an unauthenticated arbitrary-path file read vulnerability in the POST /v1/models/llm/auto-register endpoint, which accepts a caller-supplied model_path parameter without authentication or path confinement. The endpoint reads and parses config.json, tokenizer_config.json, and chat_template.jinja files at the supplied path and reflects the parsed content back to the caller, allowing an unauthenticated attacker to probe the server filesystem and extract content of files with those names in any directory.

xiaobei through 5.5.2 fails to implement authentication or signature validation on webhook endpoints, allowing unauthenticated attackers to inject arbitrary messages into the agent pipeline. Attackers can publish malicious messages via the /webhook_worktool handler and exploit unvalidated media URL fetching to perform server-side request forgery against internal services.

OGX (formerly Llama Stack, affected at commit fbe8e0f) contains an unauthenticated server-side request forgery vulnerability in the OpenAI-compatible POST /v1/responses endpoint. MCP tool definitions accept a server_url parameter (along with headers and authorization values) that is fetched server-side without destination validation; the existing validate_url_not_private() guard used for other URL inputs is not applied to server_url. On the default starter configuration, which runs without authentication, a remote unauthenticated attacker can cause the server to open connections to arbitrary internal addresses (including cloud metadata endpoints such as http://169.254.169.254/) and forward attacker-supplied headers and bearer tokens to those destinations.

Bruno versions through 4.1.0 fail to validate file paths in request body declarations, allowing attackers to read arbitrary local files by using parent-directory traversal segments. When a collection is executed, attackers can craft a request with a body:file path containing ../ sequences that resolve outside the collection directory, causing the application to read and exfiltrate arbitrary files to attacker-controlled endpoints.

Chroma 1.5.9 fails to validate maximum bounds on HNSW index parameters max_neighbors, ef_construction, and ef_search in collection-create requests. Unauthenticated attackers can supply arbitrarily large parameter values to exhaust server memory and cause denial of service during index compaction.

Aim 3.29.1 remote tracking server fails to authenticate requests and dispatches arbitrary methods through getattr without allowlist validation. Unauthenticated attackers can register clients, instantiate Repo resources, and invoke arbitrary methods to read experiments or delete runs.

Marqo 2.26.0 contains a server-side request forgery vulnerability in the add_documents endpoint that allows unauthenticated attackers to trigger requests to arbitrary URLs by supplying malicious media field values. Attackers can exploit download_image_from_url and fetch_content_sample functions which lack destination filtering and host validation to access internal services and cloud metadata endpoints.

excel-mcp-server 0.1.8 fails to enforce path confinement in stdio mode when EXCEL_FILES_PATH is unset, allowing attackers to read and write arbitrary files. Attackers can supply unchecked file paths to read and write tools to access any file accessible to the process.

cli-mcp-server 0.2.5 contains a command allowlist bypass vulnerability in the _validate_command_with_operators function when ALLOW_SHELL_OPERATORS is enabled. Attackers can use shell command substitution syntax like $(...) or backticks to execute non-allowlisted commands that bypass the ALLOWED_COMMANDS validation check.