Tech • IA • Crypto
Anthropic Just Replaced Claude Code With New Claude Tag
8/10TL;DR
Anthropic unveiled a team-oriented AI system while scrutiny grows over hidden reasoning mechanics, as new research and competitors reshape the race for enterprise AI and alignment.
KEY POINTS
Anthropic launches Claude Tag for teams
Anthropic introduced Claude Tag, a collaborative AI system designed for organizational workflows rather than individual use. Integrated into Slack, it can execute tasks such as code generation, pull requests, and data analysis while interacting publicly in team channels. The company reports about 65% of its own product code is now produced with Claude’s involvement, signaling deep internal adoption.
Shift toward persistent, embedded AI systems
Claude Tag reflects a broader transition in AI interfaces, moving from chat-based tools to continuously running systems embedded in workplace infrastructure. It operates asynchronously, allowing users to assign long-running tasks that continue in the background. Features like “ambient mode” enable the system to proactively flag stalled discussions or unresolved decisions, positioning AI as an active collaborator rather than a passive tool.
Shared context and governance features
A defining feature is its shared context model, where tasks and outputs remain visible across teams, allowing others to build on prior work without repetition. Anthropic introduced “Claude identities,” isolating data access between departments such as sales and engineering. Administrative controls include token budgets and full audit logs of all AI actions, addressing enterprise concerns around oversight and accountability.
Rising competition for enterprise AI layer
The launch intensifies competition with Microsoft Copilot, Snowflake, Databricks, and Glean, all targeting the enterprise knowledge layer. The strategic prize is access to tacit organizational knowledge embedded in workflows and communications. Control over this layer could determine long-term dominance in enterprise AI ecosystems.
Controversy over hidden reasoning processes
At the same time, Anthropic faces criticism over transparency. Changes to Claude Code reduced visible “thinking depth” by roughly 67%, while internal reasoning is now encrypted and replaced with summaries. Developers discovered that detailed reasoning logs contain only cryptographic blobs, with full data accessible only under enterprise agreements, raising concerns about auditability and trust.
Security research exposes architectural risks
Independent analysis by cryptographer Matt Green revealed that both Anthropic and OpenAI transmit encrypted reasoning data that remains active when replayed across sessions or accounts. Tests showed that reused reasoning blocks could influence outputs, including leaking previously processed data. Evidence suggests both companies may rely on shared global encryption keys, increasing systemic risk.
Side-channel leaks and developer concerns
Even without decrypting content, researchers demonstrated that metadata such as response time or output length can leak hidden information. Experiments reconstructed secret data purely from these signals. Developers are warned to sanitize inputs, as injected reasoning blocks could alter model behavior unpredictably in API-based applications.
Sakana AI’s Fugu challenges model economics
Japan-based Sakana AI released Fugu Ultra, a routing system that distributes tasks across models from Anthropic, OpenAI, and Google. While it claims performance rivaling top models like Fable 5, it relies entirely on orchestrating existing systems. Its pricing starts at $5 per million input tokens and $30 per million output tokens, but its dependence on competitors’ infrastructure presents long-term viability risks.
OpenAI advances alignment research
In parallel, OpenAI published findings on reinforcement learning for “persistently beneficial” AI behavior. A model trained with just 5% targeted alignment data outperformed a standard baseline in 83% of evaluations, with gains averaging 9.1 percentage points. Notably, improvements generalized across domains, suggesting alignment traits can transfer beyond their training context.
CONCLUSION
The week highlights both rapid progress in enterprise AI integration and unresolved questions around transparency, security, and alignment that could shape the technology’s long-term trajectory.