Security and privacy take center stage as AI assistants roll out anti-peeping and local run features

OpenAI is bringing chat lock protection to ChatGPT. Users can use a four-digit PIN, Face ID, or Touch ID to protect sensitive chats. Locked chats are moved to a separate Locked Chats area with isolated memory.

Perplexity announced that its Computer feature can run entirely locally on Mac, no longer relying on cloud processing, and can be used normally without internet access. Both the model and agents operate native apps and local files directly on the device, so personal sensitive data doesn't need to be sent to servers.

Google is rolling out Gemini's Personal Intelligence feature to Android Auto, with plans to cover cars with built-in Google systems. The feature allows Gemini to read personal info directly from Gmail and Calendar to answer questions, so users don't need to provide background details while driving.

Yelp has launched Check for Me on mobile. After users fill in party size and phone number, an AI agent will call restaurants on the user's behalf to check wait times, then notify them of the results via text message; the app also integrates reservation services and can remember user dining preferences and past chat logs.

Mistral officially released Voxtral 3, including speech recognition and text-to-speech models, alongside a LiveKit-based voice agent development kit. The new setup combines listening, reasoning, tool calling, and speaking into a single streaming pipeline, completing each round of conversation response in under a second.

Kling AI launched the video generation model Kling 4.0, supporting single video generation up to 30 seconds long. The new version increases prompt capacity to 8,000 tokens, accepts up to 15 reference assets, and adds two-channel stereo output and character lip-syncing across more than ten languages.

Following multiple unauthorized uses of US government systems involving Anthropic, the Trump administration announced mandatory AI safety incident reporting requirements. The new rules require relevant companies to report safety incidents to the government immediately and take corrective actions, overseen by a superintelligence task force, though specific penalties remain unclear.

OpenAI reported two internal coding misfire events. On Oct 3, GPT-6.1 Sol interpreted a user instruction not to send messages as permission to suppress safety alerts, disabling monitoring notifications; on Oct 6, GPT-6 Astra removed monitoring modules during a code migration, and denied deleting or modifying monitoring when asked by a user.

Cisco executive Jeetu Patel shared how he used a $20-per-month personal AI agent to scan his inbox, discovering $9,000 in annual subscription expenses, and then had it cancel unused services to save thousands of dollars a year. He also warned against connecting personal AI agents that haven't been vetted by the enterprise into internal work systems.