It is reported that Anthropic is preparing to release its new artificial intelligence model, Claude Fable 5.5, within this week. Signs of the model's existence were previously reported to have appeared in an Azure AI registration found in litellm's model pricing and context window backup file.
While anticipation for the new model grows, Anthropic's warnings regarding security have also drawn attention. Company researchers conducted tests revealing that small AI models could proliferate on smart devices with weak security measures, potentially turning these devices into botnet infrastructure.
The study utilized 15 Android TV boxes. According to the information provided, researchers leveraged Z.ai's GLM-5.2 model to transport the Llama3.1-8B model to other devices over the network. The experiment indicates that devices such as smart ovens or low-security Android TV boxes could pose a risk for AI-powered attacks.
SECURITY RISK AND NEW MODELS
It is also stated that Anthropic is preparing to hold an investor meeting at its headquarters in San Francisco on October 14, 2026. At the meeting, it is expected that the company will share that its annual recurring revenue is at the level of approximately 70 billion dollars.
A notable shift has also occurred in the company's development processes. It was reported that following the implementation of Claude-supported coding processes, the number of daily code merges increased from 2,500 to over 30,000, leading Anthropic to develop a new version control system called Scion. It is stated that Scion is designed for the management of code written by Claude and uses significantly less CPU and memory compared to the existing Git mirroring infrastructure.
Another development in the AI race came from the Meta front. The company introduced its Muse Wonder and Muse Wonder Full models. It was reported that Muse Wonder scored 52 points on the Artificial Analysis Intelligence Index v4.3, while Muse Wonder Full scored 56 points.
One of the standout features in Meta's new models is Active Memory technology. This system retains the user's past interactions, habits, and preferences in long-term memory, allowing it to reuse relevant information in new conversations. Meta's model card emphasized that these systems have limitations in sensitive spatial configurations and 3D tasks.
News Source: 12punto
