Senior AI Journalist
White House Presses OpenAI on GPT-5.6 Release, Citing Safety Concerns
The White House has reportedly asked OpenAI to “slow roll” the release of its new GPT 5.6 model due to safety concerns. This pressure comes as the Trump administration shifts from a “hands-off” approach to increased oversight of advanced AI. Instead of a broad public launch, OpenAI plans a limited preview period, with access granted “customer by customer,” potentially influenced by government evaluation.
This cautious approach mirrors Anthropic’s strategy with its Claude Mythos cyber model, which was released only to a select group of partners under “Project Glasswing.” The concern stems from the potential for powerful AI, especially cyber tools, to identify and exploit software vulnerabilities at unprecedented speeds. Such capabilities raise significant security implications for complex software infrastructure.
The administration’s request follows a recent executive order signed by Trump, directing certain AI companies to voluntarily submit new models for testing and evaluation before public release. This highlights a growing global conversation about balancing rapid AI innovation with robust safety protocols, ensuring these powerful technologies are developed and deployed responsibly.
Source: TechCrunch
AI Chatbots Exhibit Persistent Political Bias, With Gemini 3.1 Pro as a Notable Exception
A recent Washington Post investigation reveals that most major AI chatbots, including those marketed as politically neutral, consistently lean left on political questions. OpenAI’s GPT-5.5 showed the strongest bias, with 80 percent of its responses containing exclusively left-leaning arguments and only one instance of an exclusively right-leaning position. Similarly, Deepseek’s V4 Pro exhibited a 70 percent left-leaning bias in its answers.
Even xAI’s Grok 4.3, despite being promoted by Elon Musk as “truth-seeking” and “anti-woke,” provided left-leaning responses more often than right-leaning ones. This suggests that even when developers aim for neutrality or a specific political alignment, the underlying training data or subtle interventions can lead to unintended biases. The study highlights the challenge of creating truly impartial AI systems that reflect diverse viewpoints.
Google’s Gemini 3.1 Pro stood out as the most balanced model, presenting both sides of an argument 93 percent of the time. This finding indicates that it is possible to develop AI models that offer a more comprehensive and less skewed perspective on sensitive topics. The investigation underscores the critical need for continued research and development into mitigating political and other biases in AI systems.
Source: The Decoder
Google Introduces ‘Computer Use’ Capabilities to Gemini 3.5 Flash for Enhanced Agent Autonomy
Google has significantly upgraded its Gemini 3.5 Flash model by integrating “Computer Use” functionality directly into its core capabilities. This enhancement enables the AI to independently perceive, comprehend, and interact with various computing environments, including browsers, desktops, and mobile devices. Previously, such advanced capabilities were restricted to a separate Gemini 2.5 model, making this integration a major step forward for broader AI agent development.
The new “Computer Use” feature, combined with existing tools like function calls, Search, and Maps, empowers developers to construct more sophisticated and autonomous AI agents. These agents can now tackle complex tasks such as software testing automation or streamlining office workflows across different platforms. The ability for AI to interact directly with digital interfaces opens up new possibilities for efficiency and innovation in various industries.
In benchmarks, Gemini 3.5 Flash scored 78.4 on the OSWorld benchmark, outperforming Gemini 3 Flash and GPT-5.4 mini. Google also emphasizes robust safeguards against prompt injection attacks through adversarial training and optional enterprise protections, including user confirmation for sensitive actions and automatic task termination upon detecting indirect injections. This feature is now accessible via the Gemini API and the Gemini Enterprise Agent Platform, signaling Google’s commitment to making advanced agent capabilities widely available.
Source: The Decoder
What to Read Next
- Best Speculative Decoding Tools for LLMs in 2026: DSpark vs GLM 5.2 Review
- Framer vs. Durable AI: Which AI Website Builder Reigns in 2026?
- From OpenAI to SpaceX: The Race to Build Custom AI Chips Intensifies, Challenging Nvidia’s Reign
- From OpenAI to SpaceX: The Race to Build Custom AI Chips Intensifies, Challenging Nvidia’s Reign
- Browse all AI Stack Digest articles
Bookmark aistackdigest.com for daily AI tools, reviews, and workflow guides.
This article was produced with the assistance of AI tools and reviewed by the AIStackDigest editorial team.