Google’s Silent AI Install: What They’re Hiding in Your Files | Threat Wire

Google Chrome has been installing significant AI models, specifically 4GB worth of weights for Gemini Nano, onto user devices without explicit permission. This model, named `weights.bin`, powers features like scam detection and "help me write." Deleting the file results in Chrome automatically redownloading it when AI features are active, which are enabled by default. Users can only prevent this cycle by disabling Chrome's AI features via flags or enterprise policies, or by uninstalling Chrome. This behavior has been confirmed on Windows and Mac OS, with similar download mechanisms. Separately, two Linux zero-day vulnerabilities, dubbed "Dirty Frag" (CVE-202643284 and CVE-202643500), have been discovered. These kernel flaws allow local privilege escalation by chaining together an XFRMP page cache write vulnerability and an RX RPC page cache write vulnerability, enabling unauthorized kernel cache writes without system crashes or special permissions. Microsoft has confirmed their exploitation in the wild. The researcher initially found these in April but a public disclosure by a third party has prevented immediate patching or CVE assignment. In other news, the online learning platform Canvas was taken down by hackers who also stole data and demanded ransom, which Instructure, Canvas's parent company, reportedly paid. OpenAI released its ChatGPT 5.5 model, assessed to be similar to Anthropic's Mythos. Ozilla, working with Mythos, found 271 vulnerabilities in Firefox over two months. A critical vulnerability in the Google Gemini CLI (CVSS 10.0) could allow remote code execution. Google also noted thread actors using AI to develop zero-day exploits. Finally, Let's Encrypt paused certificate issuance due to an issue with cross-signed certificates.