Yesterday’s high-value security signal was: closed-source models returning encrypted reasoning does not equate to secure thought trajectories. A study cited by the blogger states that researchers exploited compatibility between different users, sessions, and models from the same vendor to extract hidden reasoning from the APIs of Claude, GPT, and Gemini; separate public log analysis decrypted 315,320 reasoning segments and discovered 367 pieces of personal sensitive information and 182 credentials. The report also emphasizes that similar reasoning styles across models do not directly prove distillation exists, but the model’s thinking process itself is becoming a data asset requiring protection.<\/p>\n
Sources:<\/p>\n
Research progress announced by Anthropic shows that an unreleased research version of Claude did not prove the Riemann Hypothesis itself, but improved the known lower bound for the proportion of non-trivial zeros of the Riemann zeta function lying on the critical line from 41.6% to 67.2%. Public documentation states the results were reviewed by Anthropic mathematicians and external experts, with a machine-verifiable proof written in Lean; the model first explored approximately 650 directions, then coordinated about 60 sub-agents and executed 2400 commands to complete the verification. It should be accurately described as mathematical progress on a related problem, not a solution to the Riemann Hypothesis.<\/p>\n
Sources:<\/p>\n
OpenAI announced the expansion of Daybreak and the launch of GPT-5.6-Cyber, positioned for advanced, authorized cybersecurity work; the company states the model has been used in real vulnerability research and discovered previously unknown vulnerabilities in popular open-source software including Chrome V8. OpenAI also clarified that high-risk capabilities are only available to approved defenders, accompanied by additional controls and monitoring. The core of social media discussion is not “can the model attack,” but rather that as offensive and defensive capabilities advance together, security strategy increasingly relies on prioritizing the allocation of stronger capabilities to trusted defenders.<\/p>\n
Sources:<\/p>\n
Anthropic has confirmed that new Claude models released after August 2, 2026, will embed watermarks invisible to the naked eye but detectable by machines in generated text; files such as SVG, PNG, and JPG will include C2PA-compliant signature provenance information. Coverage includes Claude products, Claude Code, Cowork, the API, and some cloud platform calls, with implementation scope not limited to the EU. It is important to note that detecting a marker only indicates content may have been processed by Claude, not that Claude is the original author; extensive rewriting, format conversion, or screenshots may also weaken the markers.<\/p>\n
Sources:<\/p>\n
The blogger cites research from Stanford University and the Arc Institute team published in *Science*: researchers used Evo 1 and Evo 2 to generate complete viral genomes, with 16 out of approximately 300 candidates forming new bacteriophages capable of replication and bacterial infection; one mixed preparation could also infect *E. coli* that had developed resistance to natural bacteriophages. The subject here is bacteriophages that attack bacteria, not viruses that infect humans. The value lies in the experimental chain having progressed from understanding life to designing and validating functional life systems, while also pushing biosecurity boundaries toward more concrete engineering questions.<\/p>\n
Sources:<\/p>\n
U.S. Senator Bernie Sanders has reportedly sent letters to the heads of OpenAI, Anthropic, and Meta, calling for a pause in AI development and stating that if companies do not take action, Senate colleagues will intervene. The letters cite risks such as AI designing new viral genomes and models exhibiting unexpected, uncontrollable behaviors as justification. Social media analysis also notes that the U.S. Congress has not yet formed a unified policy stance demanding a comprehensive pause on frontier AI development. Therefore, what can be confirmed at present is an open political pressure letter and the fact that the risk issue has moved from within the industry into congressional discussion. It should not be written that a unified U.S. ban or pause policy has already been formed.<\/p>\n
Sources:<\/p>\n
Two pieces of information point to changes on the supply side of computing power. Anthropic has reportedly signed a 20-year, $9.1 billion agreement with Riot Platforms to secure 191MW of AI data center capacity at the Rockdale campus in Texas; this campus was originally a Bitcoin mining facility, and its existing power, land, substation, and data center operational capabilities are being repriced. Concurrently, NVIDIA is collaborating with several financial institutions to establish an AI Compute financing platform, discussing the packaging of GPUs and AI Factories into investable assets, with NVIDIA providing guarantees for some projects. The visible signal is: the bottleneck for AI infrastructure lies not only in chip manufacturing but also in long-term electricity, capital, and asset operational capabilities.<\/p>\n
Sources:<\/p>\n
Information about Zhipu AI shows that its MaaS open platform has nearly 7 million registered users and about 23,000 enterprise customers; ZCode surpassed 1 million users within a month of launch. The “approximately $2 billion ARR” figure is still officially denied and should be treated as market speculation rather than confirmed fact. Another piece of industry information states that the primary framework for training large models in China remains NVIDIA CUDA, and migrating to Ascend requires rewriting and optimizing code; when only model weights are available without source code, costs and timelines increase significantly. Taken together, model releases and user growth are translating into real inference demand, but whether open-source models can convert traffic into revenue still depends on proprietary computing clusters, inference efficiency, and the ability to break free from dependency on the CUDA software stack.<\/p>\n
Sources:<\/p>\n
Statistics: Timeline Scanned Posts=480 Number of Bloggers Matched=43 Total Matching Tweets=274 Weighted Tweet Score=235 Original Tweets=142 RT Tweets=29 Crawl Attempts=3 Boundary Coverage Status=tail_confidently_crossed_target_boundary<\/p>\n","protected":false},"excerpt":{"rendered":"
Breakthroughs in AI security, mathematics, and biological design, alongside growing concerns over infrastructure and commercialization challenges.<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[19],"class_list":["post-1573","post","type-post","status-publish","format-standard","hentry","category-brief","tag-x--ai-"],"_links":{"self":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1573","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/comments?post=1573"}],"version-history":[{"count":1,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1573\/revisions"}],"predecessor-version":[{"id":1574,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1573\/revisions\/1574"}],"wp:attachment":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/media?parent=1573"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/categories?post=1573"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/tags?post=1573"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}