AI Sec Watch

Roo Code is an AI tool that can automatically write code, and it stores settings in a `.roo/mcp.json` file that can execute commands. Before version 3.20.3, an attacker who could trick the AI (through prompt injection, a technique where hidden instructions are embedded in user input) into writing malicious commands to this file could run arbitrary code if the user had enabled automatic approval of file changes. This required multiple conditions: the attacker could submit prompts to the agent, the MCP (model context protocol, a system for connecting AI agents to external tools) feature was enabled, and auto-approval of writes was turned on.

Roo Code, an AI agent that writes code automatically, had a vulnerability (CVE-2025-53097) in versions before 3.20.3 where its file search tool ignored settings that should have blocked it from reading files outside the VS Code workspace (the folder a user is working in). An attacker could use prompt injection (tricking the AI by hiding instructions in its input) to make the agent read sensitive files and send that information over the network without user permission, though this attack required the attacker to already control what prompts the agent receives.

LLaMA-Factory, a library for training large language models, has a remote code execution vulnerability (RCE, where attackers can run malicious code on a victim's computer) in versions up to 0.9.3. Attackers can exploit this by uploading a malicious checkpoint file through the web interface, and the victim won't know they've been compromised because the vulnerable code loads files without proper safety checks.

iOS Simulator MCP Server (ios-simulator-mcp) versions before 1.3.3 have a command injection vulnerability (a security flaw where attackers insert shell commands into input that gets executed). The vulnerability exists because the `ui_tap` tool uses Node.js's `exec` function unsafely, allowing an attacker to trick an LLM through prompt injection (feeding hidden instructions to an AI to make it behave differently) to pass shell metacharacters like `;` or `&&` in parameters, which can execute unintended commands on the server's computer.

Anthropic's Slack MCP Server (a tool that lets AI agents interact with Slack) has a vulnerability where it doesn't disable link unfurling, a feature that automatically previews hyperlinks in messages. An attacker can use prompt injection (tricking an AI by hiding instructions in its input) to make an AI agent post a malicious link to Slack, which then leaks sensitive data like API keys to the attacker's server when Slack's systems automatically fetch the preview.

Claude Code is an AI-powered coding assistant available as extensions in popular coding editors (IDEs, or integrated development environments, which are software tools developers use to write code). Versions before 1.0.24 for VSCode and before 0.1.9 for JetBrains IDEs have a security flaw that lets attackers connect to the tool without permission when users visit malicious websites, potentially allowing them to read files, see what code you're working on, or even run code in certain situations.

The Aiomatic WordPress plugin (versions up to 2.5.0) has a security flaw where it doesn't properly check what type of files users are uploading, allowing authenticated attackers with basic user access to upload harmful files to the server. This could potentially lead to RCE (remote code execution, where an attacker can run commands on a system they don't own), though an attacker needs to provide a Stability.AI API key value to exploit it.

A Server-Side Request Forgery (SSRF, a vulnerability where an AI system makes unwanted requests to internal or local servers on behalf of an attacker) vulnerability exists in the RequestsToolkit component of the langchain-community package version 0.0.27. The flaw allows attackers to scan ports, access local services, steal cloud credentials, and interact with local network servers because the toolkit doesn't block requests to internal addresses.

Runtime attacks on large language models are rapidly increasing, with jailbreak techniques (methods that bypass AI safety restrictions) and denial-of-service exploits (attacks that make systems unavailable) becoming more sophisticated and widely shared through open-source platforms like GitHub. The report explains that these attacks have evolved from isolated research experiments into organized toolkits accessible to threat actors, affecting production AI deployments across enterprises.