English Summary:

OpenAI is expanding its Trusted Access for Cyber (TAC) program to thousands of verified defenders and hundreds of teams responsible for critical software defense. This initiative introduces GPT-5.4-Cyber, a variant of GPT-5.4 specifically fine-tuned for cybersecurity use cases, which lowers the refusal boundary for legitimate cybersecurity work. This enables advanced defensive workflows, including binary reverse engineering capabilities.

OpenAI emphasizes an approach of scaling cyber defense in lockstep with increasing AI model capabilities, guided by three principles: democratized access, iterative deployment, and ecosystem resilience. The company acknowledges that cyber risk is already present and accelerating, and aims to help defenders find and fix vulnerabilities faster. Tools like Codex Security have already contributed to fixing over 3,000 critical vulnerabilities.

Access to TAC is granted through identity verification for individuals at chatgpt.com/cyber and via OpenAI representatives for enterprises. This effort underscores the continuous need for evolving defenses to ensure AI safety, especially in preparation for future, more powerful models.

English Summary:

Anthropic has launched Claude Managed Agents, a new platform designed to simplify AI agent deployment for enterprises. This platform aims to embed orchestration logic directly into the AI model layer, promising to reduce agent deployment time from weeks or months to just days. It eliminates the need for complex aspects like sandboxing code execution, checkpointing, and credential management.

However, this approach raises concerns about increased vendor lock-in, as it shifts more control over enterprise AI agent deployments and operations to Anthropic. Storing session data in an Anthropic-managed database increases the risk of enterprises becoming tied to a single vendor's system. This could make guaranteeing agent behavior more challenging, especially for highly sensitive workflows such as financial analysis or customer-facing tasks.

The pricing model is a hybrid of token-based billing and a usage-based runtime fee, charging $0.08 per hour when agents are actively running. This is noted as being less predictable compared to competitors like Microsoft's Copilot Studio and OpenAI's Agents SDK. While Claude Managed Agents offers reduced engineering overhead and faster deployment, it presents enterprises with a trade-off: ease of use versus potential loss of control, observability, portability, and the risk of vendor lock-in.

English Summary:

A Chinese AI system, based on a "double agent" approach developed by researchers, has solved and verified an open conjecture in commutative algebra from 2014 in just 80 hours. This system divides the mathematical process into two functions traditionally performed by humans: one agent for informal reasoning, exploring hypotheses and building proof structures, and a second agent for formal verification, translating proposals into rigorous, machine-checkable mathematical language.

This achievement signifies a major shift not just in faster computation, but in how AI can participate in scientific discovery, particularly in mathematics. Human intervention was limited to providing access to restricted documents the system couldn't obtain on its own, with no direct mathematical supervision or process correction. This redefines the human role from problem-solving to facilitating information access and validating final results.

However, the work has not yet undergone peer review, necessitating validation from the scientific community. This advancement fits into a broader context of China's intensified investment in AI, suggesting that mathematical research is becoming a new arena for technological and geopolitical competition. The system demonstrates that the entire process, from initial intuition to final verification, can be partially automated, raising profound questions about mathematical research, education, and the role of human thought in scientific discovery.