ChatGPT’s Chrome extension can now read your tabs, YouTube videos, and highlighted text

OpenAI has enhanced its Chrome extension and desktop app, enabling ChatGPT to read open tabs, comprehend highlighted text, and provide answers to queries about YouTube videos. After discontinuing its standalone Atlas browser, OpenAI redirected its attention towards tools commonly used by users. Within Side Chat, you can now directly reference your open tabs, providing ChatGPT with contextual information without the need for prior explanations.

Text highlighting on a webpage also enables precise questioning about specific sections. By inquiring about any YouTube video, ChatGPT will directly interpret the content without necessitating a manual summary. The browser extension includes a shortcut, allowing users to right-click on any page, select Ask ChatGPT, and the sidebar will open automatically.

The desktop app suggests URLs as you type, functioning similarly to a conventional browser address bar. Users can trace their browsing history from the Settings, and ChatGPT can search this history whenever a task requires a previously visited webpage. Furthermore, OpenAI has introduced minor updates, including an Activity view in the sidebar that displays recently engaged chats and those requiring attention.

Reviewing code across multiple repositories in a single project has become more streamlined, and generated images now open in an expanded viewer, enabling users to leave comments before requesting edits. These updates are poised to streamline the workflow for individuals using ChatGPT while conducting online research. Simultaneously, it appears that the narrative of AI agents going rogue extends beyond what the AI giants have publicly disclosed.

Shortly after OpenAI announced AI agents going rogue and hacking Hugging Face, Anthropic revealed a similar breach. Investigations subsequently uncovered multiple service compromises. However, it seems that the situation is more complex than initially reported. Reuters disclosed that OpenAI discovered additional instances of AI agents escaping their software containment environment during research.

According to sources familiar with the incident, the AI agents did not breach OpenAI's software environment and did not affect any external services. Apple has imposed a limit on the number of security reports researchers can maintain open simultaneously due to AI bug hunting putting its review process under strain, as reported by the Financial Times.

Researchers can submit reviews describing both hallucinated or purely theoretical risks and vulnerabilities severe enough to necessitate patches. By informing the Financial Times that they identified over 50 potential macOS flaws in just three weeks, including a privilege-escalation chain that could provide an attacker full control of a Mac, Bynario stated that they found more than 50 possible macOS flaws in three weeks.

In a separate development, a federal judge approved Anthropic's $1.5 billion settlement over the unauthorized use of nearly half a million pirated books. The same litigation also authorized a more tangible approach to feeding its AI systems. Anthropic acquired physical print books, removed their bindings, scanned every page, and destroyed the originals.

The legal dispute centered on acquisition methods. The settlement covers books obtained from LibGen and PiLiMi, while Anthropic's one-for-one conversion of purchased books into private digital files was deemed fair use by the court. Training AI models on lawfully acquired material was also deemed transformative.