According to CoinWorld, Anthropic has launched an agent identity security mechanism for its team collaboration product, Claude Tag, aiming to completely block backdoors for AI data leaks. This mechanism equips each channel with an independent permission package, directly assigning dedicated accounts to AIs to prevent unauthorized access in multi-person collaborations. The new security architecture replaces the traditional borrowing of personal credentials with dynamic policy configuration files, enabling AIs to function as independent team-level identities. Security sandbox expert Kenton Varda criticized the model, arguing that directly assigning dedicated accounts to AIs is unsuitable for large-scale tasks and undermines human accountability mechanisms. He pointed out that agents cannot be legally and administratively held accountable; all operational permissions must originate from a real, living person. Varda advocates for a capability-based security model to ensure that AI actions are traceable to the specific initiator.

In the past hour, total liquidations across the network reached $10.09 million, with $8.53 million from long positions, with $1.55 million from short positions. BTC liquidation $2.39 million, ETH liquidation $3.43 million.

According to CoinWorld, OpenAI's ChatGPT subscription and quota system experienced a widespread anomaly, causing confusion regarding usage quotas and model permissions for users at different levels. Some Plus subscribers suddenly displayed a 20x limit, gaining access to the higher quotas previously reserved for Pro subscribers. Simultaneously, some Pro subscribers paying $200 per month found their usage quotas abnormally reset to 75%, or their quota limits increased by four times. The system malfunction also affected model usage permissions; many Pro subscribers reported that while their account subscription status displayed Pro, their actual performance had been downgraded to Plus level, and the GPT-5.5 Pro flagship inference model was unavailable in the model selection list. To date, OpenAI's official status page only indicates the issue may occur when users are using Codex access tokens, stating that mitigation measures have been implemented. The official stance has not been publicly acknowledged or explained regarding the quota calculation and account level synchronization anomalies experienced by ChatGPT subscribers. The community generally believes that a faulty backend fee calculation and synchronization logic is the underlying cause of the rapid depletion of quotas and abnormal rate limiting.

According to CoinWorld, Notion has announced a public beta of its external agents platform as part of its business and enterprise initiatives. This platform supports integrating Anthropic's Claude and AI programming tool Cursor into team collaboration dashboards. Users can @ agents or directly assign tasks within the Notion workspace, allowing AI to automatically collaborate on code writing, data analysis, and content generation. The newly beta Claude agent runs on the Anthropic Managed Agent (CMA) architecture. Users do not need to configure API keys and are paid for per run using Notion credits. After task assignment, the model can autonomously read documents and execute multi-step planning. However, it's important to note that the CMA architecture operates in a stateful manner; Notion's Zero Data Residue (ZDR) commitment does not apply to the Claude agent, and session data will remain on the server. Notion has also opened an external agent interface, allowing users to bind their local Cursor agent to the Notion workspace for zero Notion credit consumption and automated development.

According to CoinWorld, GLM-5.2, a model developed by Zhipu AI, has received official ARC-AGI benchmark validation, demonstrating performance comparable to OpenAI's lightweight GPT-5.5. Data shows that GLM-5.2 achieved an accuracy of 22.8% on the ARC-AGI-2 benchmark, with an average run cost of $0.25 per task. On the slightly less challenging ARC-AGI-1 benchmark, GLM-5.2 achieved an accuracy of 77.0%, with a run cost of $0.19 per task. ARC Prize, the benchmarking organization, stated that testing of GLM-5.2 on the more challenging ARC-AGI-3 benchmark is currently suspended due to limitations in inference throughput; the team plans to release specific test results only after a reliable evaluation is possible.