Anthropic’s Claude Is Becoming Part of the Research Team

The most consequential disclosure in today’s AI news came from Anthropic, which says Claude now leads 26% of the company’s AI research and development—up from below 1% in March—while humans still supervise the work. That claim is arriving alongside fresh evidence of the industry’s unresolved liabilities: contested training data, exploitable AI infrastructure and demands for genuinely independent safety evaluation.

Claude’s Role in Anthropic’s R&D Is Expanding Fast

Anthropic says Claude can complete most tasks end-to-end from a high-level prompt while remaining under human supervision. The company says more than 90% of its R&D is done in collaboration with Claude, but stresses that the model is not fully autonomous for any measured subset of the work. Business Insider reports on Anthropic’s figures and their rise from March.

Anthropic is presenting the metrics as a way to track progress toward recursive self-improvement, which it defines as a model autonomously building its successor. It also says that models accelerating their own development could make them harder for humans to understand or control, and is calling on other labs to publish comparable measurements. The immediate implication is narrower but important: frontier labs are beginning to quantify how much of their own development is being delegated to their models.

Unsealed Documents Reopen the Publisher Conflict

Newly unsealed court documents show Microsoft and OpenAI employees discussing whether AI training based on millions of news articles could damage the publishing industry. The reporting includes a Microsoft executive’s description of OpenAI’s web scraping as the “largest theft of labor in human history,” but the evidence reflects internal concerns reported from court filings rather than a new ruling. The New York Times details the discussions.

The significance is that the dispute is not only about whether publishers’ material was used. The documents, as described by the reporting, show executives recognizing a possible threat to the content ecosystem that AI systems rely on. That makes the commercial and legal question more circular: the tools may depend on sources whose economics they could weaken.

Huawei Pulls Its Next AI Chip Forward

Huawei is accelerating the launch of its Ascend 960DT AI chip to the first quarter of 2027, with an Ascend 960PR variant planned for the third quarter, according to Bloomberg reporting. The company is positioning the chips as part of a broader effort to replace Nvidia in China. Bloomberg reports the revised schedule.

Huawei’s rotating chairman also expects a major domestic shift toward Ascend-based systems for model training in 2027, despite supply constraints, according to the South China Morning Post. The company’s challenge is broader than a single processor: its strategy includes networking and clustered systems designed to connect large numbers of chips. Claims that the new chip will deliver twice the performance of its predecessor remain company claims, not an independently verified comparison with Nvidia hardware.

A Claude-Assisted Attack Exposed OpenAI Weaknesses

Hacktron, a cybersecurity startup, says it used Claude in authorized research to exploit a vulnerability affecting OpenAI’s community help forum and gain access to an employee’s ChatGPT and Codex accounts. The team said it stopped before accessing internal code, disclosed the issue to OpenAI and received a $6,500 bounty. Business Insider describes the disclosure and the attack path.

The episode is a concrete example of the dual-use problem around capable models. Anthropic’s Cyber Verification Program relaxed certain cyber restrictions for the authorized research, while the researchers used the model to probe another frontier lab’s systems. It does not establish that Claude independently carried out the intrusion, but it does show how model assistance can compress the work needed to find and exploit weaknesses in AI companies’ own infrastructure.

Safety Evaluation Faces a Test of Independence

A group of AI evaluation organizations has published proposed minimum conditions for third-party assessors working with labs such as OpenAI and Anthropic. The group says credible evaluators need scientific objectivity, transparency, independence and protection from interference, along with access to relevant systems, data, tools and physical spaces. Business Insider reports on the letter backed by Geoffrey Hinton and Stuart Russell.

The proposal follows commitments from OpenAI and Anthropic leaders to welcome outside evaluators, but it highlights the difference between access offered by a lab and oversight that can operate without the lab controlling the terms. The evaluators are also asking for unfiltered communication with boards and other oversight bodies, and for the ability to release findings publicly. Those conditions will determine whether “independent evaluation” is an accountability mechanism or an extension of internal review.

SoftBank Adds Credit as Its AI Expansion Continues

SoftBank secured an additional $450 million, increasing an existing credit line to $6.5 billion, people familiar with the matter told Bloomberg. The financing gives the Japanese conglomerate more flexibility as it expands its AI investments, while the report frames the move against rising AI-related debt. Bloomberg reports the increased credit line.

This is not evidence by itself of financial distress, but it is a useful counterpoint to the industry’s capability narrative. The race to build and deploy AI is also a financing race, and the scale of credit available to major investors will shape how quickly ambitious projects can proceed—and how much balance-sheet risk accompanies them.

The Next Accountability Test Is Operational

Today’s developments connect capability to the institutions meant to contain its risks. Anthropic is measuring how much of its R&D Claude can lead; Huawei is accelerating the hardware race; and a Claude-assisted security test found a route into OpenAI accounts. The near-term test is whether labs can turn calls for transparency into arrangements that outsiders can verify—while resolving the publisher and liability disputes created by the systems they are scaling.