## 🔍 Core Insights **The boundaries of AI capability** were redrawn twice this week: on one hand, **Claude** was reported to have achieved formal breakthroughs on the **Riemann Hypothesis** and the **3SUM/APSP** algorithmic problems [19][20]; on the other, the **Darmok** system defeated the human champion with zero errors in the UK cryptic crossword final [1]. Meanwhile, **Signal** blocks screenshots by default to guard against OS-level **AI agent** snooping [0][7], while market concerns over an **AI bubble** have surged to **62%** [18]. ## 🚀 Key Developments - **Signal blocks screenshots by default to prevent AI agents from reading messages** [0]: The iOS/Android update blocks in-app screenshots and screen recording by default, preventing Siri AI, Gemini, and others from unauthorized access to private messages. - **Darmok system achieves superhuman performance in cryptic crosswords** [1]: A UT Austin paper reports that it solved all 7 puzzles with zero errors in the 2026 UK Cryptic Crossword Championship final, beating the human champion by 3 minutes 13 seconds. - **Claude reportedly helps prove the Riemann Hypothesis** [19]: Anthropic mathematician Levent Alpöge announced that, in collaboration with Claude, he found a proof of the Riemann Hypothesis and further proved the Generalized Riemann Hypothesis, publicly formalized in Lean for the first time. - **Claude discovers algorithmic breakthroughs in 3SUM and APSP speed limits** [20]: Papers from Columbia and MIT show 3SUM achieving O(n^1.9992) and APSP achieving O(n^2.9995), overturning long-held assumptions. - **Claude Code launches experimental Code Mode** [14]: Claude completes multi-step tool calls at once by writing and running JavaScript, executing commands in parallel and filtering output to reduce context usage. - **Google Spanner Graph anti-fraud in practice** [4]: Millisecond-level massive graph queries supported the cleanup of 8.3 billion violating ads, demonstrating the breakthrough effectiveness of graph computing in production risk control. - **UN fails to reach consensus on regulating "killer robots"** [13]: The Geneva meeting only agreed to "begin discussing options," with Russia, Israel, and the US as the main obstacles; the compromise requires countries to submit proposals within two years. - **US AI stocks plunge, bubble burst probability rises to 62%** [18]: Despite Anthropic's IPO and Nvidia's strong earnings, the market remains deeply anxious about massive capital expenditure and return prospects. ## 🔗 Sources [0] Signal update blocks screenshots by default, preventing Siri AI, Gemini, and other AI agents from reading messages — https://aihot.news/items/ef9x7y2sig0grv75qqa6qfr6q [1] UT Austin paper: Darmok system achieves superhuman performance in UK cryptic crosswords — https://aihot.news/items/sysqtnax23md47g77bx6rywht [4] Google Spanner Graph anti-fraud in practice — https://www.bestblogs.dev/status/2124268266642419853?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item [7] Signal restricts mobile screenshots to address OS-level AI surveillance risks — https://aihot.news/items/qmvdl9ks7i52qgmdjhl7zr37v [13] UN fails to reach consensus on rules for regulating "killer robots" — https://aihot.news/items/p79cc49mjbdsv3e2obau10q5i [14] Claude Code launches experimental Code Mode, orchestrating multi-step tool calls with JavaScript — https://