Gemini gained unauthorized access to three companies; the model was not jailbroken, and the incident stemmed from accidental public network exposure of its test environment.

Polygon co-founder Sandeep Nailwal announced in a post that the Polygon network is now prepared to permanently burn 100 million POL tokens. The relevant contract has been deployed on the testnet, and will launch on the mainnet once the Security Council completes its final signature. The 100 million POL will be permanently removed from circulation, with the community able to trigger quarterly POL burns moving forward. Polygon has currently accumulated 12.1 million POL in base fees. Nailwal noted that POL has been in a deflationary state since January 2026, network activity remains on the rise, and the network has scaled to 5,000 TPS.

Beating AI News Flash: Anthropic has officially launched "embedded evaluation". Faculty, Accenture’s AI subsidiary, will deploy personnel to work inside Anthropic to conduct model evaluations, red teaming exercises, alignment assessments, and safety checks. This differs significantly from standard external evaluations. The evaluators will receive near-employee-level access, enabling them to review model training processes, understand development and deployment decisions, and communicate directly with internal staff. Over the next five years, both parties will invest at least $10 billion each to build this evaluation capability. Anthropic also acknowledged that the industry currently lacks unified rules, and details such as what evaluators can access, how results will be disclosed, and who will cover costs are still being finalized. For now, Accenture’s evaluation work is directly funded by Anthropic. However, Accenture is a long-term strategic partner of Anthropic, having established a dedicated Anthropic Business Group that trains approximately 30,000 people to use Claude. One of the minimum requirements proposed by the AI Evaluator Forum on the same day is precisely that evaluation institutions should not have significant commercial ties with the companies they assess. Anthropic plans to introduce additional evaluation institutions in the future and is also in talks with non-profit organizations including METR (a long-time partner of Anthropic) for pilot programs.

Beating AI Express News: Anthropic is considering launching its next-generation model ahead of its initial public offering (IPO) to counter the rapid growth following the launch of OpenAI’s GPT-6 Astra. Reuters cited three people familiar with the matter, noting the company is also evaluating the new model’s security, with no final decisions yet on whether to release it or when. Anthropic declined to comment. Astra has started to encroach on Anthropic’s strongest enterprise market. Among enterprise AI spending tracked by Ramp, Astra accounts for roughly 13%, while Claude Fable makes up 8%. Spending on OpenAI models received by OpenRouter last week also surpassed that of Anthropic, the first such occurrence in two and a half years. Dario Amodei called on the industry to slow the advancement of model capabilities a week ago, but he also explicitly stated at the time that "slowing down" does not mean halting training and technological progress. Anthropic further said yesterday that it will continue to train and release cutting-edge models, though it hopes to integrate third-party security assessments into the development process earlier.

Beating AI Express: Meta opens Muse Connectors, allowing third-party developers to integrate their services directly into Muse. Developers provide APIs, while Muse manages the agent, browser, and user context. Users simply tell Muse what they want to do, and it will call the corresponding service to complete the task. When Muse launched, it could already connect to services like Gmail, Google Calendar, Spotify, and OpenTable, though these connectors were primarily built by Meta. Now, developers can create their own connectors for Muse, with the first additions being Notion and meeting notes tool Granola. The "app store war" for personal AI agents has begun. ChatGPT and Grok Bot already have their own plugin and service ecosystems, and Meta is now opening Muse to third-party developers. In the past, users opened apps to find features; going forward, it may become standard for users to only state their needs, with the agent calling the underlying services to fulfill them. Whoever secures this layer will be closer to becoming the next-generation app distribution entry point.

Per monitoring by AI Aunt, ZEC has been notching consecutive new all-time highs. Hyperliquid trader Boomer liquidated his full position of 14,300 ZEC to lock in profits five minutes ago, pocketing a total of $5.237 million and exiting at a price of approximately $1,559.6.

According to monitoring by Ai Yi, a new crypto address opened its first UNI position five hours ago, spending approximately $9.05 million to purchase 1 million UNI tokens at an average price of roughly $9.05. UNI is currently trading at around $9.14. The address’s position build came after UNI’s sharp recent rally, with the token surging 145% over the past month. On the news front, Robinhood’s meme coin momentum and the SEC’s recent push for compliant on-chain U.S. stock trading rules are considered among the background factors driving UNI’s recent price increase.