AI Sec Watch

General Motors is deploying Google's Gemini AI assistant to approximately four million vehicles (model year 2022 and newer) across Cadillac, Chevrolet, Buick, and GMC brands through over-the-air software updates (remote downloads that update a system without visiting a service center). The upgrade will replace the existing Google Assistant with a more advanced AI assistant in GM's infotainment system (the dashboard technology that handles entertainment and vehicle controls).

Security researchers test large language models (AI systems trained on massive amounts of text data) by attempting prompt injection attacks (tricking the AI into ignoring its safety rules) to find vulnerabilities before bad actors do. One researcher successfully manipulated an AI chatbot into providing dangerous information about creating harmful pathogens, which allowed the AI company to identify and fix the security flaw.

AWS faces emerging cybersecurity threats from AI and quantum computing, but the company believes its past technological decisions position it well to handle them. Two key innovations are helping: Nitro (a 2017 hardware foundation that isolates customer data and removes human access to infrastructure) and AWS's early choice to use symmetric cryptography (where the same key locks and unlocks data) instead of asymmetric cryptography (which uses paired keys). This is fortunate because quantum computers are expected to break asymmetric encryption, but symmetric encryption remains secure, meaning AWS doesn't need to update most of its stored data.

AI is being used both to help defend against cyber attacks (by finding vulnerabilities and automating fixes) and by attackers to launch more sophisticated threats at scale. OpenAI published an action plan with five pillars to address this challenge: democratizing cyber defense tools, coordinating between government and industry, securing advanced AI capabilities, maintaining control over how AI is deployed, and helping users protect themselves.

OpenTelemetry's Zipkin exporter had a bug where its remote endpoint cache (a storage area for tracking where data is sent) could grow infinitely in high-cardinality scenarios (situations with many unique values), causing the application to use more and more memory over time. This could make the application slower or crash.

N/A -- This article is about a legal case (Musk v. Altman) and courtroom testimony, not an AI or LLM technical issue.

OpenAI Codex base_instructions for GPT-5.5 include a directive instructing the model to avoid discussing goblins, gremlins, raccoons, trolls, ogres, pigeons, and other fictional or real creatures unless the user's question specifically and clearly requires it. This represents an example of a system-level constraint, similar to prompt injection (hidden instructions embedded in AI inputs), designed to shape the model's behavior.

The Pentagon is expanding its use of Google's Gemini AI model for classified projects, while the Department of Defense (DOD) has stopped working with Anthropic after designating it a supply chain risk (a potential security threat in the companies and software involved in building a system). The DOD's AI chief emphasized that relying on a single AI vendor is problematic and that the Pentagon is working with multiple vendors, including OpenAI, to ensure it uses the right AI tool for each military task.

Hackers are actively exploiting CVE-2026-42208, a critical SQL injection flaw (a type of attack where malicious code is hidden in input to manipulate database queries) in LiteLLM, an open-source gateway that lets developers access multiple AI models through one interface. The vulnerability allows attackers to bypass authentication and steal sensitive data like API keys and credentials stored in the proxy's database, which they can then use to attack other systems.