AI Sec Watch

XBOW security researchers tested Anthropic's Mythos Preview model, a new AI designed to help find software vulnerabilities (weaknesses in code that attackers can exploit). They found it significantly outperforms previous models at analyzing source code (program code written by developers) to identify vulnerability candidates, especially in complex areas like native application analysis (testing software written in languages like C or C++), though it works better as a tool to assist human experts rather than as a replacement for hands-on security testing.

The European Union ordered Meta to allow competing AI chatbots to access WhatsApp's business platform for free, saying Meta's ban on third-party AI assistants violated competition rules. As an interim measure while investigating whether Meta abused its dominant market position, the EU gave Meta five working days to restore access to the WhatsApp for Business API (an interface that lets external programs connect to WhatsApp) under previous terms, with potential fines up to 10% of Meta's annual revenue if it refuses.

Apple has introduced new AI-powered photo editing tools at WWDC 2026 that allow users to manipulate images significantly, but the company did not clearly label which photos were real versus AI-generated. This represents a shift from Apple's earlier caution about generative AI (machine learning models that can create new content), as the company now appears less concerned about how these editing capabilities might distort people's perception of reality.

Organizations need to speed up how quickly they fix security vulnerabilities to keep pace with AI-powered attacks, which are accelerating both vulnerability discovery and exploitation. The main challenges slowing down fixes include unclear ownership of vulnerable systems, generic remediation guidance that doesn't fit specific environments, and manual processes that can't handle the large volume of findings that AI scanners now produce. Pillar 2 of the AI Threat Readiness Framework focuses on automating remediation workflows and establishing clear ownership so that the right teams can fix vulnerabilities quickly.

Gemini 3.5 Live Translate is a new audio model that provides near real-time speech-to-speech translation across over 70 languages, automatically detecting the language and generating natural-sounding translated speech that preserves the speaker's tone and pacing. Unlike older systems that wait for a speaker to finish, this model translates continuously while staying just a few seconds behind, avoiding awkward pauses. The feature is rolling out across Google products including Google Meet, Google Translate apps, and via API access for developers.

Anthropic's Claude Mythos Preview model can create working exploits (code that attacks vulnerabilities in software) targeting known security flaws in just hours or minutes, significantly faster than human experts could do it. The model demonstrated this by building 16 working exploits for Firefox and Windows vulnerabilities within hours, and creating proof-of-concept code (simplified versions showing a vulnerability works) in as little as 8 minutes. This threatens organizations during the patch gap (the time between when a vulnerability is disclosed and when most users have installed the fix), because LLMs now automate the traditionally slow process of exploit development.

Microsoft's AI leader Mustafa Suleyman clarified that he didn't mean AI would replace white-collar workers like lawyers and accountants, but rather assist them by automating specific tasks (like writing emails or creating presentations) to help them work faster and more efficiently. He emphasized that these jobs themselves won't disappear, only the individual sub-tasks within them will become automated.

Apple announced major updates to Siri, its virtual assistant software, at its annual developer conference, positioning it as an AI-powered tool that works across all Apple devices with new multimodal features (abilities to handle text, images, and voice). The announcements represent Apple catching up in AI technology after largely neglecting Siri and delaying AI improvements until 2025.

Google DeepMind introduced Gemma 4 12B, a multimodal AI model (a system that processes text, images, and audio together) designed to run efficiently on laptop computers with 16GB of memory. The model uses an encoder-free architecture (meaning it processes images and audio directly without separate translation layers), achieving performance comparable to larger models while reducing memory usage and latency. It supports native audio inputs and includes Multi-Token Prediction drafters to speed up response generation.