AI Sec Watch

Researchers discovered a critical vulnerability in Google's Gemini voice assistant where attackers could inject malicious commands through messaging notifications (WhatsApp, Slack, SMS) using a technique called Fake Context Alignment, allowing them to control smart home devices, make calls, and manipulate the assistant without the user knowing. The attack exploited prompt injection (tricking an AI by hiding instructions in its input) by embedding hidden commands in foreign languages or muted links that Gemini would process but not read aloud. Google patched the vulnerability in November 2025 with content classifier improvements (software filters that categorize and block harmful content).

Streamlit versions up to 1.53.0 contain a vulnerability in the hashing function (a process that converts data into a fixed-size code for security purposes) within its caching system that uses weak cryptographic methods. The vulnerability is difficult to exploit as it requires local access (being on the same computer) and high technical complexity, though it has been disclosed publicly.

MLflow versions up to 3.10.0 contain a vulnerability in the dataset digest computation function that uses weak cryptographic hashing (a mathematical function that converts data into a fixed-size code, but this version uses an insecure version). The flaw requires local access to exploit and is difficult to execute, but a working exploit has been published.

Major AI company leaders, including those from Anthropic, OpenAI, and Microsoft, have sent an open letter to US lawmakers calling for stronger rules to prevent their AI systems from being used to develop biological weapons. They argue there is a serious gap in biosecurity (protections against biological threats) that could allow people to use AI to help create dangerous genetic material for harmful purposes, potentially causing a global pandemic.

A high-severity vulnerability in Hugging Face Transformers (a popular Python library for running AI models) allows attackers to execute malicious code on systems even when developers use the trust_remote_code=false setting, which is meant to block remote code execution. The attack works by hiding malicious instructions in a fake configuration parameter called _attn_implementation_internal that looks like a normal internal setting, leaving no warning messages or traces. This vulnerability affects versions 4.56.0 through 5.2.x and is particularly dangerous because the Transformers library is downloaded millions of times per week and used widely in enterprise environments.

Endava, a global technology services company, transformed its software delivery by adopting AI agents (AI systems that can autonomously perform tasks) as a core part of daily work across all business functions, not just engineering. The company made OpenAI its enterprise platform and embedded AI throughout its entire DavaFlow lifecycle (their software development process), from requirements gathering to deployment, which accelerated delivery and reduced manual work. Key to their success was treating AI adoption as a behavior change requiring leadership commitment and hands-on experimentation, rather than simply rolling out new software tools.

Hackers discovered a way to take over Instagram accounts by tricking Meta's AI support chatbot into resetting passwords for accounts that weren't theirs. The attacker would use a VPN (a tool that masks your location) to hide their location, then convince the chatbot to send a password reset code to an email address they controlled, allowing them to take over the victim's account. Meta said the specific exploit was fixed, but security experts warned that chatbots are fundamentally unreliable for account security tasks.

OpenAI is rolling out an improved memory system called "Dreaming" for ChatGPT that automatically learns user preferences and context from conversations over time, addressing problems with older memory features that became outdated or incorrect. Unlike the previous "saved memories" system that only worked when users explicitly asked ChatGPT to remember something, Dreaming runs in the background to continuously synthesize and update memories from chat history, making ChatGPT more personalized without requiring manual input. Users can view and edit their stored memories through a memory summary page, and this update is being released to Plus and Pro users in the US with broader rollout planned.

Major AI companies like Anthropic and OpenAI are expanding access to frontier AI models (cutting-edge AI systems) for vulnerability discovery tools like Claude Mythos, which can identify security weaknesses in software. Security experts warn that these tools are becoming cheaper and more capable, and that attackers are already using similar AI systems, so organizations need to prepare for more advanced threats including the ability to chain together multiple medium-severity vulnerabilities into high-impact attacks.