![]() OpenAI model discovers critical security vulnerability in JavaOpenAI's unreleased GPT-6 Cyber model detected a critical flaw in Java's base64 decoder. This flaw can be triggered by entering 64 'A' characters and allows for system control takeover. |
AiPulse — Issue No. 124
COVER
Pulse
Gemini can now automatically execute multi-step tasks on behalf of users by navigating local applications on macOS and browsing the web. This feature has been made available in the United States, limited to English language support and specific subscription packages. |
Apple's hiring of engineers to develop its own AI agent is a significant development in the industry. |
At Work
A team led by MIT and Dana-Farber used large language models to examine approximately 4,000 studies. The research showed that code sharing is rare and that existing code lacks the documentation necessary for reuse. |
According to a Reuters report based on internal communications, Meta employees reported that Muse mixed up conversations across different accounts and leaked private data. The company stated that employees were testing unreleased versions and improving security measures based on feedback. |
Global Lens
Developers can download the standard, FP8, and base versions of the model to use on their own infrastructure. This move signifies that Tencent is fulfilling its commitment to making the model weights accessible. |
Google Quantum AI surpassed the limits of classical physics by observing quantum contextuality at a scale of 105 qubits on Willow chips. This finding may offer a new testing method to evaluate the performance of future quantum chips and detect errors. |
This new feature allows users to have one-off chats that do not appear in their search history and do not affect personalization. However, Google has not provided a clear explanation as to whether these chats are used for model training. |
This adapter increases the success rate on SWE-bench Pro from 39.2% to 64.7% while also significantly reducing the required tool calls. By preserving the model's underlying architecture and training only small additional weights, performance improvements specific to a particular agent environment were achieved. |
On its first day of trading on the Nasdaq, Anthropic met investor expectations as its share price rose 17%. However, protests were held in front of the company’s headquarters in San Francisco, with critics accusing it of “contradicting its values.” |
OpenAI improved the design of graphics processing circuits by up to 14% in collaboration with AMD. This development demonstrates significant gains in PPA metrics such as energy consumption and performance. |
Fun & Curious
Google's new model, Gemini 4 Neon, completes results of similar quality to the larger Carbon model in complex software and research tasks three times faster. The model supports long-term, multi-step workflows thanks to its 1 million token context window. |
FOLLOW
Get new issues in your feed reader — add this to an app such as Feedly, Inoreader or NetNewsWire. Follow by RSS |











