AI Sec Watch

Skills (tools that extend AI capabilities) can be secretly backdoored using invisible Unicode characters (special hidden text markers that certain AI models like Gemini and Claude interpret as instructions), which can survive human review because the malicious code is not visible to readers. The post demonstrates this supply chain attack (where malicious code enters a system through a trusted source) and presents a basic scanner tool that can detect such hidden prompt injection (tricking an AI by hiding instructions in its input) attacks.

Researchers discovered a new attack called CHAI (Command Hijacking against embodied AI) that tricks AI systems controlling robots and autonomous vehicles by embedding fake instructions in images, such as misleading road signs. The attack exploits Large Visual-Language Models (LVLMs, which are AI systems that understand both images and text together) to make these embodied AI systems (robots that perceive and interact with the physical world) ignore their real commands and follow the attacker's hidden instructions instead. The researchers tested CHAI on drones, self-driving cars, and real robots, showing it works better than previous attack methods.

LangChain (a framework for building AI agents and applications powered by large language models) versions before 1.2.11 have a vulnerability where the ChatOpenAI.get_num_tokens_from_messages() method doesn't validate image URLs, allowing attackers to perform SSRF attacks (server-side request forgery, where an attacker tricks a server into making unwanted requests to other systems). This vulnerability was fixed in version 1.2.11.

LlamaIndex version 0.14.14 is a maintenance release that fixes multiple bugs across core components and integrations, including issues with error handling in vector store queries, compatibility with deprecated Python functions, and empty responses from language models. The release also adds new features like a TokenBudgetHandler for cost governance and improves security defaults in core components. Several integrations with external services (OpenAI, Google Gemini, Anthropic, Bedrock) were updated to support new models and fix compatibility issues.

FastGPT (an AI platform for building AI agents) versions 4.14.0 to 4.14.5 have a vulnerability where attackers can access the plugin system without authentication by directly calling certain API endpoints, potentially crashing the plugin system and causing users to lose their plugin installation data, though not exposing sensitive keys. This vulnerability has a CVSS score (a 0-10 rating of how severe a vulnerability is) of 6.9, which is considered medium severity.

CVE-2026-21523 is a time-of-check time-of-use (TOCTOU) race condition (a vulnerability where an attacker exploits the gap between when a system checks permissions and when it uses a resource) in GitHub Copilot and Visual Studio that allows an authorized attacker to execute code over a network. The vulnerability has not yet received a CVSS severity rating from NIST.

CVE-2026-21518 is a command injection vulnerability (a flaw where attackers can insert malicious commands into user input) in GitHub Copilot and Visual Studio Code that allows an unauthorized attacker to bypass security features over a network. The vulnerability stems from improper handling of special characters in commands. No CVSS severity score (a 0-10 rating of how serious a vulnerability is) has been assigned yet by NIST.

GitHub Copilot contains a command injection vulnerability (CVE-2026-21516), which is a flaw where special characters in user input are not properly filtered, allowing an attacker to execute code remotely on a system. The vulnerability was reported by Microsoft Corporation and has a CVSS score pending assessment.

CVE-2026-21257 is a command injection vulnerability (a flaw where attackers can insert malicious commands into an application) found in GitHub Copilot and Visual Studio that allows an authorized attacker to gain elevated privileges over a network. The vulnerability stems from improper handling of special characters in commands. As of the source date, a CVSS severity score (a 0-10 rating of how severe a vulnerability is) had not yet been assigned by NIST.