Security vulnerabilities, privacy incidents, safety concerns, and policy updates affecting LLMs and AI agents.
Ollama before version 0.1.46 has a security flaw where attackers can use path traversal (a technique that manipulates file paths to access files outside their intended directory) in the api/push route to discover which files exist on the server. This allows an attacker to learn information about the server's file system that should be private.
Fix: Update Ollama to version 0.1.46 or later.
NVD/CVE DatabaseOllama before version 0.1.34 has a vulnerability where the CreateModelHandler function improperly reads user-controlled file paths without limits, allowing an attacker to specify a blocking file like /dev/random, which causes a goroutine (a lightweight process in Go) to run infinitely and consume resources even after the user cancels their request. This is a resource exhaustion (CWE-404: Improper Resource Shutdown or Release) issue that can disrupt service availability.
A vulnerability in Ollama before version 0.1.46 allows an attacker to crash the application by uploading a malformed GGUF file (a model format file) using two HTTP requests and then referencing it in a custom Modelfile. This causes a segmentation fault (a type of crash where the program tries to access memory it shouldn't), making the application unavailable.
Ollama versions through 0.3.14 have a vulnerability where the api/create endpoint leaks information about which files exist on the server. When someone calls the CreateModel route with a path that doesn't exist, the server returns an error message saying 'File does not exist', which allows attackers to probe the server's file system.
A vulnerability in langchain version 0.2.5's GraphCypherQAChain class allows attackers to use prompt injection (tricking an AI by hiding instructions in its input) to perform SQL injection attacks on databases. This can let attackers steal data, delete information, disrupt services, or access data they shouldn't have access to, especially in systems serving multiple users.
CVE-2024-7774 is a path traversal vulnerability (a security flaw where attackers can access files outside the intended directory) in langchain-ai/langchainjs version 0.2.5 that allows attackers to save, overwrite, read, and delete files anywhere on a system. The vulnerability exists in the `getFullPath` method and related functions because they do not properly filter or validate user input before handling file paths.
A vulnerability exists in the GraphCypherQAChain class of langchain-ai/langchainjs versions 0.2.5 that allows prompt injection (tricking an AI by hiding instructions in its input), which can lead to SQL injection (inserting malicious database commands). This vulnerability could allow attackers to manipulate data, steal sensitive information, delete data to cause service outages, or breach security in systems serving multiple users.
CVE-2024-48142 is a prompt injection vulnerability (a technique where attackers hide malicious instructions in text sent to an AI) in Monica ChatGPT AI Assistant v2.4.0 that lets attackers steal all chat messages between a user and the AI through a specially crafted message.
A prompt injection vulnerability (tricking an AI by hiding instructions in its input) was found in Monica Your AI Copilot v6.3.0, a ChatGPT-powered browser extension. Attackers can exploit this flaw by sending a specially crafted message to access and steal all chat data between the user and the AI assistant, both from past conversations and future ones.
CVE-2024-48145 is a prompt injection vulnerability (a type of attack where malicious instructions are hidden in text input to an AI system) in Netangular Technologies ChatNet AI Version v1.0 that allows attackers to steal all chat data between users and the AI by sending a specially crafted message. The vulnerability is classified under CWE-77 (improper neutralization of special elements used in commands), meaning the system fails to properly filter dangerous input before processing it.
CVE-2024-48144 is a prompt injection vulnerability (tricking an AI by hiding instructions in its input) in Fusion Chat Chat AI Assistant Ask Me Anything v1.2.4.0 that allows attackers to craft a malicious message in the chatbox to steal all previous and future conversations between the user and the AI assistant. The vulnerability is caused by improper handling of special elements in user input (CWE-77, a weakness in command injection prevention).
CVE-2024-48141 is a prompt injection vulnerability (a technique where an attacker hides malicious instructions in text sent to an AI) in Zhipu AI CodeGeeX version 2.17.0's chatbox. An attacker can craft a message to trick the AI into leaking all previous and future chat conversations between the user and the assistant.
CVE-2024-48139 is a prompt injection vulnerability (a technique where attackers hide malicious instructions in messages sent to an AI) in Blackbox AI version 1.3.95 that allows attackers to steal all chat messages between a user and the AI by sending a specially crafted message. This vulnerability is classified as a command injection flaw (where attackers manipulate input to execute unintended commands).
Cursor is a code editor that uses AI to help with programming. Before September 27, 2024, attackers could trick Cursor's command generation feature into running harmful commands if a user imported a malicious website into the prompt and the attacker used prompt injection (hidden instructions in text that manipulate AI behavior) on that website. A server-side patch was released quickly to block dangerous characters, and Cursor version 0.42 added client-side protections and a new preview box setting that requires manual approval before commands run.
CVE-2024-49361 is a vulnerability in ACON, a machine learning library that performs adaptive correlation optimization. The vulnerability exists in how ACON validates input data, which could allow an attacker to bypass these checks and execute arbitrary code (run commands they shouldn't be able to run) on systems using ACON. Machine learning applications that accept user-provided data are at the highest risk, especially those running on production servers (live systems serving real users).
Gradio, an open-source Python package for building user interfaces, has a cross-site scripting vulnerability (XSS, where malicious code hidden in files runs in users' browsers) that affects servers allowing file uploads. Attackers can upload harmful HTML, JavaScript, or SVG files that execute when other users view or download them, potentially stealing data or compromising accounts.
Gradio, an open-source Python package for building demos, has a vulnerability where the connection between the FRP client and server (fast reverse proxy, a tool that exposes local applications to the internet) isn't encrypted when the `share=True` option is used. This means attackers can intercept and read files uploaded to the server or modify data being sent, putting sensitive information at risk for users sharing Gradio demos publicly online.
Gradio, an open-source Python package for building AI demos, has a race condition (a bug where two operations interfere with each other due to timing) in its configuration function that lets attackers change the backend URL. This could redirect users to a fake server to steal login credentials or uploaded files, especially affecting Gradio servers accessible over the internet.
Fix: Update Ollama to version 0.1.34 or later.
NVD/CVE DatabaseFix: Update Ollama to version 0.1.46 or later.
NVD/CVE DatabaseLangflow v1.0.12 contains a remote code execution vulnerability (RCE, where an attacker can run commands on a system they don't own) in its PythonCodeTool component. This flaw allows attackers to execute arbitrary code through the tool. The vulnerability was publicly disclosed in October 2024.
PyTorch versions 2.4.1 and earlier contain a vulnerability in RemoteModule that allows RCE (remote code execution, where an attacker can run commands on a system they don't own) through deserialization of untrusted data. However, multiple parties dispute whether this is actually a security flaw, arguing it is intended behavior in PyTorch's distributed computing features (tools for running AI computations across multiple machines).
Fix: A patch is available at https://github.com/langchain-ai/langchainjs/commit/a0fad77d6b569e5872bd4a9d33be0c0785e538a9
NVD/CVE DatabaseFix: A server-side patch released on September 27, 2024 prevents newlines or control characters from being streamed back. Cursor 0.42 includes client-side mitigations that block newlines or control characters from entering the terminal directly. Users can enable the setting `"cursor.terminal.usePreviewBox"` and set it to `true` to stream responses into a preview box that must be manually accepted before inserting into the terminal. The patch is applied server-side, so no additional action is needed on older versions. Additionally, Cursor's maintainers recommend only including trusted context in prompts as a best practice.
NVD/CVE DatabaseFix: Upgrade to gradio>=5. As a workaround, restrict uploads to non-executable file types (like images or text) and implement server-side validation to sanitize or reject HTML, JavaScript, and SVG files before they are stored or displayed to users.
NVD/CVE DatabaseFix: Users should upgrade to `gradio>=5` to fix this issue. As an alternative, users can avoid using `share=True` in production environments and instead host their Gradio applications on servers with HTTPS enabled (a secure protocol that encrypts communication) to ensure safe data transmission.
NVD/CVE DatabaseFix: Upgrade to gradio>=5 (version 5 or newer). The source notes there are no known workarounds for this issue.
NVD/CVE Database