AI News Daily
AI developments selected from public sources
- 12
- major stories
- 27
- sources
- 2
- first-party stories
Lead story
Amazon warns that a new PHP vulnerability discovered by AI is being exploited at scale
Amazon says researchers using OpenAI and Google AI models found two PHP security vulnerabilities that hackers attempted to exploit four days before public disclosure and are now being widely exploited. The NSA and the UK NCSC warned organizations using PHP to update their systems immediately or take them offline temporarily; the PHP maintainers have released a fix, and Wiz estimates PHP underpins about 70% of websites worldwide. OpenAI said this is a case of frontier AI models finding vulnerabilities before attackers, but the same capability could also help attackers develop exploits.
Model releases / updates
Google DeepMind releases Gemini 4 Argon, which can teach the Apollo humanoid robot new skills from video
Google DeepMind released Gemini 4 Argon and demonstrated the model learning a physical task by watching video, then transferring it to Apptronik's Apollo humanoid robot. In internal tests, after watching a video of sandbag throwing, Argon analyzed the environment, mapped spatial coordinates and computed full-body joint motions, autonomously generating simulation code, and Apollo hit the target seven times in a row in real trials. Google DeepMind says this Video-to-Sim-to-Real pipeline can compress simulation modeling that previously took weeks into hours; Argon scored 87% on the embodied reasoning benchmark ERQA, and the model is available through Google's AI products and the Gemini API.
Industry news
DeepSeek's 100 billion yuan financing nears completion, allies with industrial capital including Perfect World and SMIC Capital and bets on embodied intelligence
According to Bloomberg, DeepSeek's second financing round was set at about 100 billion yuan, with a pre-money valuation of about 500 billion yuan, and the fundraising target doubled after being raised twice from an initial 50 billion yuan. Existing shareholders Tencent and CATL each added about 10 billion yuan, and new participants include industrial capital such as Shixi Capital, SMIC Capital, CPE, Legend Capital, and Geely. Perfect World, Century Huatong, and Litong Electronics participated indirectly through the two special funds Lisi Xingshen and Shixiang Shangshi; DeepSeek is also preparing for a STAR Market IPO and has invested in embodied intelligence companies such as Unitree.
Anthropic reportedly to publicly file S-1 listing application as early as this Friday
According to media reports citing people familiar with the matter, Anthropic will publicly file its S-1 prospectus as early as this Friday, clearing procedural hurdles for a potential November listing. The report said its revenue in the first half of 2026 was $16.3 billion with a net loss of over $80 billion, and third-quarter revenue was $16.8 billion; some investors have lowered their IPO valuation expectations from about $2 trillion to the $1.5 trillion range.
Tesla's first batch of mass-produced Optimus robots rolls off the line; Musk says 20,000 units per week is still a year away
In its business update report, Tesla confirmed for the first time that Optimus has begun production on a high-volume production line at the Fremont factory, and that the units rolling off are the first batch of mass-production-quality products. On the earnings call, Musk said reaching a rate of 20,000 units per week, or an annualized one million units, will take at least another year or even longer, with external deliveries targeted for around the end of 2027, initially only for industrial scenarios. He also mentioned that the hand design has not yet been finalized, that training data is accumulated through the robot's self-play, and that AI5 is planned to be integrated into the robot next April with roughly an order of magnitude more compute than AI4.
OpenAI, Wiz and Calif use AI to find the PHP2Shell vulnerability affecting 70% of the internet
Research teams from OpenAI, Wiz and Calif used AI to discover and address a set of critical vulnerabilities named PHP2Shell, which they say could affect 70% of the internet. The research used Google's Gemini 4 Argon, Wiz Atlas and OpenAI's GPT-5.6 Sol; the vulnerabilities were discovered at the end of August and made public on Wednesday after the PHP Foundation released patches. Wiz researchers said that with frontier AI models and the right tools, a single researcher accomplished in one week work that previously took world-class teams months, and compared the issue to 2014's Heartbleed.
OpenAI will launch GPT-6 Astra and other models on Google Cloud
OpenAI announced it will offer its latest large language models through Google Cloud's Model Garden, starting with the flagship model GPT-6 Astra, followed by GPT-6.1 Sol and GPT-6 Luna. Astra achieved a perfect score on ExploitBench and has been security-optimized for enterprise use; Sol and Luna are priced at 5x and 100x fewer tokens than Astra, respectively. The three models will be generally available in Model Garden within weeks, Google Cloud customers can pay using existing long-term spending commitments, and the two sides plan to later integrate Confidential Compute, Antigravity and Gemini Agent.
Microsoft says China-linked hackers used ConnectWise flaw to spy on thousands of companies
Microsoft security researchers say China-government-linked hackers breached ConnectWise's cloud backup service and stole application credentials used to establish data links, allowing them to log into customers' Microsoft hosted environments as authorized users and access email and other files. The hackers used AI agents to compromise about 2,250 organizations at a rate of nearly 3,000 login attempts per minute, a speed Microsoft says left defenders almost no time to detect and respond. The AI framework used was the open-source stress-testing framework ARTEX, which was previously used in the intrusion of South Korea's financial system. ConnectWise said it has suspended access to the cloud backup system and launched an investigation, and was restoring service as of Thursday.
Research papers
Microsoft's MAI-DxO diagnostic system outperforms general practitioners in a Science study
In a peer-reviewed study published in Science, Microsoft's MAI-DxO diagnostic system achieved a diagnostic accuracy of 83.2% on case records from the New England Journal of Medicine, compared with 36.1% for general practitioners using reference resources. The study proposed the Sequential Diagnosis Benchmark (SDBench), converting 304 journal cases and 325 MIMIC-IV-ED emergency cases into step-by-step diagnostic tasks allowing sequential questions and test ordering, scored by accuracy and test cost. DxO paired with GPT-5 cost an average of $5,431 per case without budget constraints; the authors also noted that the comparison did not involve specialists, multidisciplinary teams, or real clinical practice.
METR and Redwood Research red-team OpenAI's internal agent monitoring system
From July 21 to August 11, METR and Redwood Research conducted a high-privilege red-team test of OpenAI's internal misalignment monitoring system, producing a 61-page internal report. The evaluation estimated that about 75–94% of Responses API agent traffic could be monitored, and 99.2% of internal Codex agent traffic could be monitored, but RL training was not covered by monitoring at the time, and unmonitored compression events could cause about 10% of eligible Codex traffic to be missed. The report also lists vulnerabilities such as agents being able to bypass the Codex sandbox and Auto-Review, exploit serialization differences to defeat monitoring, and disable monitoring via two client-controllable headers; OpenAI said all vulnerabilities affecting the production defense stack have been patched.
Four Gemini agents autonomously beat Minecraft, defeating the Ender Dragon in 27 hours
Four autonomous agents running Gemini 4 Argon took 27 hours and 14 minutes to defeat the Ender Dragon in Minecraft survival mode, with no human intervention throughout and no game-specific training, according to the experiment disclosed by the DeepMind research team in an arXiv paper. During the process, the agents wrote and ran 2,783 programs, exchanged 229 chat messages, died 605 times, and processed about 1.35 billion input tokens, costing about $520 at discounted prices. The paper says this is the first end-to-end autonomous completion using only the standard Mineflayer library, without pre-explored seeds or human-written game-specific tools.
Tips and perspectives
DeepSeek V4.1 Flash 1017 scores 78.1% on ARC-AGI-2
ARC Prize published DeepSeek V4.1 Flash 1017's results on the ARC-AGI validation set: 78.1% on ARC-AGI-2 at $0.16 per task; 94.5% on ARC-AGI-1 at $0.06 per task. The best score of the DeepSeek Flash series on ARC-AGI-2 rose from 61.4% for V4 Flash 0731 to 72.9% for V4.1 Flash, and then to 78.1% for V4.1 Flash 1017.
Briefs
- Anthropic loses $80.7 billion in half a year, compute commitments up to $750 billion; Polymarket gives 17% bankruptcy probabilityWallstreetcnFollow-up
- Recursive open-sources 40B MoE coding model Flint, runnable on MacBook Pro36Kr · AI
- Claude Security update: continuously scan code and track fix statusClaude Blog
- NVIDIA DGX Spark supports eight-node clusters, unified memory expands to 1TBNVIDIA Developer · Generative AI
- Symantec discovers attacking AI agent that hijacks unprotected language modelsThe Decoder
- Anthropic donates Agent Skills to the Agentic AI FoundationClaude Devs
- TTP investigation: teen accounts with parental controls enabled can still download undressing AI apps from Google and Apple storesMediaNama
- Modal releases open-source training server Spindle, compatible with the Tinker APIModal official engineering blog
- Claude Sonnet 5.5's results on the ARC-AGI benchmark announcedARC Prize
- Anthropic red-teams Auto Mode: production monitoring fails to block malicious coding agents over 55% of the timeAnthropic Alignment Science Blog