AI Daily · Nov 21, 2026 · SatAI Daily
ICREAT AI NEWS
Thinking Machines releases open-weight model Inkling 2 Mini
Thinking Machines released the open-weight recurrent Transformer model Inkling 2 Mini, which scored 20 on the Artificial Analysis Intelligence Index.
Product releases/updates
browser-use 1.0 replaces 24 Agent tools with a single Python tool
browser-use 1.0 was released on November 10, 2026, replacing the 24 browser actions of 0.x with one main tool, python, a persistent REPL with access to the raw Chrome DevTools Protocol.
Industry news
Anthropic completes largest IPO ever, valued at $1.5 trillion
Anthropic completed an $81.3 billion IPO at $83 per share, surpassing SpaceX to become the largest IPO ever; $ANT closed at $97.17 on its first trading day, up 17% from the offer price.
Signal blocks screenshots on iOS 27 and Android to prevent AI agents from reading private chats
Signal blocks screenshots on iOS 27 and, on supported Android versions, triggers protection when an AI assistant attempts to access the screen, to prevent AI agents from reading private chat content. Signal told Reuters that privacy faces an "existential threat," noting that system-level AI such as Apple's Siri AI and Google's Gemini Intelligence can capture screen content to complete tasks, thereby bypassing end-to-end encryption.
Cloudflare explains the five nginx vulnerabilities in nginx2root and origin server protection scope
F5 and nginx.org disclosed ten nginx vulnerabilities on November 18, among which nginx2root, reported by the Xint team, includes five memory corruption vulnerabilities that unauthenticated attackers can use to crash nginx workers and, according to their research, execute commands.
Bloomberg analyzes ten million job postings: AI-exposed roles in China decline about 10 percentage points faster
A Bloomberg analysis found that in the three years after June 2022, total online job postings in China halved, with roles highly overlapping with AI capabilities declining about 10 percentage points faster than other roles. Chen Qin of ShanghaiTech University analyzed 116 million job postings and estimated that AI caused a net 5.2% reduction in new hiring from 2023 to early 2026; a study involving the Chinese Academy of Social Sciences found a 21% decline in online positions and 4% lower average wages.
OpenAI discloses adversarial distillation defense method: GPT-6.2 Astra reasoning trace extraction attack success rate drops to 0.2%
OpenAI published detection and blocking measures for adversarial distillation, combining adversarial testing, model training, classifiers, and API-layer protections, and strengthening protection of models' protected reasoning (chain-of-thought).
Anthropic's record IPO fails to halt AI selloff, global stocks fall
After Anthropic completed a record IPO, AI-related selling continued in global stock markets, with Asian stocks posting their worst week since 2020. Anthropic shares rose nearly 40% at one point on Thursday and closed up 17%, then fell as much as 10% in Friday premarket trading, slipping back near the $83 offer price; Taiwan stocks fell 6.7%, TSMC fell about 6%, the Nikkei fell 5%, South Korea's Kospi fell 5.5%, and SoftBank fell 14.5% in Tokyo.
Paper research
Researcher plants 2 million canary tokens in the AI supply chain, finding privacy leaks
Security researcher HD Moore planted 2 million digital canary tokens in prompts sent to AI services to track how sensitive information flows through the AI supply chain, and found extensive leaks in this ecosystem.
Testing 65 AIs with 127,000 questions: their advice to everyone is highly similar
The author asked 65 AIs from 30 companies everyday questions a total of 127,918 times, finding that 73% of book recommendations pointed to the same book, 98% of investment answers were index funds, and 90% of profession answers were computer science. On the open-source model OLMo, which can be measured stage by stage, this preference appeared at the supervised fine-tuning step, while after pretraining it was still close to random. The author's suggested mitigations are to provide specific personal details or to have the model list ten options and choose yourself; the former reduced the probability of matching answers from 53% to 11%.
Tips and opinions
Levent Alpöge says he discovered a proof of the Riemann hypothesis in collaboration with Claude
In alpo.ge/remark.pdf, Levent Alpöge says that in collaboration with Claude he discovered a proof of the Riemann hypothesis, that the method generalizes to the generalized Riemann hypothesis, and that he expects it to further cover the extended Riemann hypothesis. He says the two have formalized the argument without assumptions and describes Claude as far surpassing him near the finish line. Simon Willison shared this statement in a quoted post.
Anthropic shares fall below offer price on first trading day, AI stocks shed over 50% in five months
Gary Marcus says Anthropic's IPO is heading toward the most embarrassing ever: ordinary investors saw an opening price of 111 on the first trading day, a close down to 97, and a further premarket drop of about 9% to around 88. He notes the decline spread across the whole market, with SpaceX expected to open below its $135 offer price, SoftBank down 14.5% overnight, and Asian markets also hit. He argues that in this environment OpenAI cannot IPO at a $1 trillion valuation and may run out of cash.
Briefs
- Google Gemini for Mac adds Computer Use, can operate desktop apps in the backgroundThe Decoder
- Google previews Semantic governance policies on the Gemini platform, intercepting tool calls before agents execute themGoogle Developers Blog
- METR releases Q4 2026 capability assessment: AI R&D-related capabilities improving rapidly and becoming harder to measureMETR Blog
- Hugging Face releases harness adapters: using RL to adapt models to a specified Agent harnessHugging Face Blog
- Baseten releases decision models bison-1 and groundhog-1 based on Gemma 4Baseten Blog