Anthropic Just Replaced Claude Code With New Claude Tag

1. Anthropic's Claude Tag: AI for Team Collaboration

Timestamp: 00:00:18 to 00:02:32 - watch this moment on skim

Anthropic has launched Claude Tag, an evolution of Claude Code, designed for team-wide collaboration rather than individual developers. It integrates into platforms like Slack, breaking down requests, executing tasks with connected tools, and posting results publicly within team channels. This aims to shift collaboration from private chats to transparent, shared contexts where progress and reasoning are visible to all, enabling asynchronous work and proactive surfacing of issues.

Significance (High): This represents a significant shift in how AI is integrated into the workplace, moving from individual assistance to collective team augmentation. The emphasis on shared context and asynchronous execution could dramatically improve team productivity and transparency.

Sources in support: Speaker (Analyst), Andrej Karpathy (AI Researcher at Anthropic), Rob Seeman (General Manager at Slack)

2. The Mystery of Anthropic's 'Extended Thinking'

Timestamp: 00:03:43 to 00:06:58 - watch this moment on skim

Concerns have surfaced regarding Anthropic's 'extended thinking' feature in Claude Code, where users observed a significant decrease in thinking depth after default settings were changed. Instead of detailed reasoning, logs showed encrypted cryptographic signatures. This practice, where the actual reasoning is encrypted and held by Anthropic, means users only receive a summary, with full access requiring an enterprise agreement, leading to accusations of data loss and opacity.

Significance (High): This lack of transparency in AI reasoning processes erodes user trust and raises questions about the true capabilities and limitations of the models. The move to encrypt core reasoning processes, accessible only under specific commercial agreements, suggests a potential conflict between user understanding and corporate control.

Sources in support: Speaker (Analyst), Patrick McKenna (Developer)

Neutral sources: Anthropic (AI Safety and Research Company)

3. Cryptographic Analysis of AI Reasoning Blocks

Timestamp: 00:07:08 to 00:09:59 - watch this moment on skim

Cryptography professor Matt Green's analysis of reasoning blocks from OpenAI and Anthropic revealed they are sent as base-64 encoded ciphertext. While designed to carry state in stateless conversations, the implementation shows vulnerabilities. Both providers appear to use a single global encryption key, allowing for replay attacks across different sessions and accounts. Furthermore, side-channel attacks using block length or response time can leak information, even without decrypting the data.

Significance (High): These findings expose significant security and privacy risks in current AI model implementations. The potential for data leakage and manipulation through replay or side-channel attacks suggests that the current methods for handling AI reasoning state are inadequate for sensitive applications.

Sources in support: Speaker (Analyst), Matt Green (Cryptography Professor at Johns Hopkins)

Neutral sources: OpenAI (AI Research and Deployment Company), Anthropic (AI Safety and Research Company)

4. Sakana AI's Fugu: A Dispatcher Model

Timestamp: 00:11:21 to 00:13:07 - watch this moment on skim

Sakana AI has released the Fugu model series, with Fugu Ultra claiming to rival top-tier models like Fable 5. However, Fugu is not a base model but a dispatcher that routes tasks to specialized models like Claude Opus 4.8 or GPT 5.5 based on task classification. While its routing logic is sophisticated, its performance benchmarks reflect the aggregated capabilities of the underlying frontier models, raising questions about its long-term business viability against direct competition.

Significance (Medium): This 'selling water' model highlights a trend of AI orchestration layers. While offering flexibility, its dependence on foundational model providers creates a structural risk, as those providers could eventually view it as competition, potentially disrupting its service.

Sources in support: Speaker (Analyst), Sakana AI (AI Research Lab)

Neutral sources: OpenAI (AI Research and Deployment Company), Anthropic (AI Safety and Research Company)

Encryption Key Management Concerns (High severity)

Type: Professional

Researchers Matt Green and Patrick McKenna found that OpenAI and Anthropic may be using single global encryption keys for client-side reasoning data, potentially compromising security and privacy. This practice, coupled with issues in replay protection and side-channel leakage, raises questions about the robustness of their AI security protocols.

Significance: This raises profound questions about the security and trustworthiness of AI models. If reasoning states are not properly secured per session or account, sensitive data could be exposed or manipulated. The implications for enterprise adoption and user trust are immense, suggesting a critical need for improved key management practices.