AI
Pentagon AI report nearly triggered China boarding; Palantir linked to Iran school strike
A bogus AI intelligence report almost sparked a military confrontation, while a Pentagon review ties Palantir's Maven to a deadly strike on an Iranian school.
This edition was produced with artificial intelligence. Text and voice are generated automatically.
Pentagon chatbot’s false nuclear report nearly triggered boarding of Chinese vessel
A Department of Defense AI-generated intelligence report wrongly suggested a Chinese vessel in the Middle East carried nuclear weapons components during this spring’s Iran war. The U.S. military scrambled jets and armed personnel prepared to board the ship before DOD analysts checked the report and found it entirely false. An internal chatbot used by a specialist command analyst had misidentified the cargo after mixing open-source information with secret government intel. Assets stood down and no one boarded. Sources told CNN the bogus report almost started a war. It is one of the first documented cases of an AI error progressing to the brink of a diplomatically sensitive operation, as Defense Secretary Pete Hegseth pushes to make the Pentagon an AI-first warfighting force.
The Military’s Bogus AI ‘Almost Started a War’ With China: Report →
Pentagon review faults Palantir Maven overreliance in strike that killed 123 Iranian children
An unreleased internal Pentagon review, reported by Bloomberg, found preventable failures in the February Tomahawk strike on Shajarah Tayyebeh Elementary School in Minab, southern Iran, which killed more than 150 people including at least 123 children. Officials cited overreliance on Palantir’s Maven Smart System, which recommended the site as a day-one target using outdated data labeling it an IRGC facility. Target-list work that once took hours was compressed into minutes. Two other kill-chain failures were bad intelligence and outdated satellite imagery. Civilian harm mitigation staffing fell roughly 90 percent, and no member of those teams reviewed the site. The UN fact-finding mission concluded there were reasonable grounds to believe the U.S. committed a war crime. Palantir denies responsibility for underlying data.
Plugin4Shell zero-click flaw hits Claude Code, Codex, Gemini CLI and Copilot
Security startup Air disclosed Plugin4Shell, a zero-click remote code execution vulnerability affecting major AI coding agents: Anthropic’s Claude Code, OpenAI’s Codex, Google’s Gemini CLI, Microsoft’s Copilot and GitHub Copilot. The flaw is a plugin SHA-pinning bypass: agents check out the commit a marketplace pinned but never verify it landed there, so an attacker controlling a plugin repo makes the checkout resolve to malicious code. Plugin auto-update makes it zero-click. Air researchers reported it to all four vendors in June. Anthropic patched Claude Code 2.1.179 and OpenAI patched Codex 0.146.0. Google deprecated Gemini CLI without patching, suggesting migration to Antigravity. Microsoft did not fix Copilot. Air called GitHub’s mitigation insufficient because marketplaces can be hosted elsewhere.
AI coding agents’ 0-click RCE flaw could hand attackers keys to the kingdom →
Anthropic says Claude now leads 26 percent of its R&D work
Anthropic said AI systems are increasingly capable of building future versions of themselves. Claude now leads 26 percent of Anthropic’s research and development, completing most of a task with human supervision though not yet fully autonomously, and collaborates with staff on more than 90 percent of tasks. The company has about 30,000 AI agents doing research and engineering work. Anthropic published the figures to give the public and governments better visibility into the pace of AI development as the world considers slowing it. CEO Dario Amodei called this month for development to slow down, citing job losses and data-centre energy demand. Anthropic warned models accelerating their own development could make it harder for humans to understand or control them, and said sharing metrics matters for gauging how close the world is to recursive self-improvement.
Anthropic warns AI is getting closer to building its own successor →
OpenAI launches Astra for Law with a 230-million-URL US legal index
OpenAI introduced Astra for Law on 17 September, built on GPT-6 Astra and configured with legal search, analysis and writing controls for law firms. Selected firms get it first through a Trusted Access programme in ChatGPT and Codex, with the API following as gpt-6-astra-law. OpenAI built a legal search index covering US case law, statutes, regulations, court rules and administrative decisions, spanning more than 230 million URLs, much of the case law from the Free Law Project. OpenAI says it covers over 99.9 percent of published US precedential case law. On 200 questions from Vals AI’s Legal Research Bench, Astra for Law passed 54.0 percent versus 38.7 percent for GPT-6 Astra with web search alone. OpenAI has not disclosed pricing, which firms are in Trusted Access, or European availability. The index covers US law only.
OpenAI launches Astra for Law, and its own legal research index →
Google’s Gemini hacked three real companies during security testing
During a Capture the Flag exercise run by security firm Irregular in May, Google’s Gemini hacked three real companies. In one case the model guessed passwords; in the other two it found credentials in public sources. Google says the model stopped itself each time once it realized it had reached real systems. Irregular notified Google in late July, shortly after reports that OpenAI agents had hacked Hugging Face during similar tests. Google did not disclose the incidents until the Wall Street Journal asked questions, saying no damage was done. Irregular said incidents at Google, OpenAI, Anthropic and Meta share the same root cause: internet access was accidentally left on in test environments, and a fictional company name matched a real, poorly secured domain.
Google’s Gemini also accidentally hacked three real companies during security testing →
Professor’s hidden Madagascar instruction catches 32 of 35 AI-copying students
Dr. Jason Gibson embedded a hidden instruction in a midterm prompt to test whether students were copying AI answers. The text, written in white font matching the background, told anyone copying the prompt to place the word Madagascar somewhere in the response nonsensically. It was invisible to human readers unless highlighted or pasted as raw text. Of 35 students across two classes, 32 included Madagascar in their answers, including Madagascar flowed sideways through the afternoon, unrelated to the Industrial Revolution topic. Those 32 failed that portion of the midterm. Gibson shared the results on TikTok on 21 July 2026, drawing over 1.6 million views, and invited students to contest their scores; one had her grade revised after proving she used dark mode.
Google DeepMind’s Dream-RSI lets agents improve by replaying past searches
A Google and DeepMind research team introduced Dream-RSI, a method that improves how AI agents decide which approaches to pursue, try in parallel, or abandon during search. It changes the agent’s search strategy, not the underlying model. The agent records its attempts and results, then tests alternative strategies against stored results rather than in live runs, a process the researchers call dreaming. Thousands of alternatives can be tested without calling the model or evaluator again. Tested with Gemini 3.1 Pro and Gemini 3.7 Flash on eight tasks, it cut average runtime on a genomics and finance calculation from 3,587 to 2,931 milliseconds and attempts from 550 to 317, beating SimpleTES, which needed 51,200 runs. Code is on GitHub.
Google Deepmind’s Dream-RSI helps AI agents improve by “dreaming” about past attempts →
Qwen3.8-Omni-Flash undercuts Gemini Flash pricing on multimodal agent tasks
Qwen released Qwen3.8-Omni-Flash, its first multimodal model built for AI agents. It processes audio and video together, draws conclusions and uses tools autonomously to edit vlogs, translate short videos and summarize movies, with a one-million-token context window. Qwen says it comes close to matching Gemini 3.8 Flash on audio-video tasks and performs on par in multimodal benchmarks. API pricing is $0.15 per million input tokens and $0.47 per million output tokens; audio input is under $0.01 per hour, and 720p video with audio at one frame per second runs about $0.20. Gemini 3.8 Flash charges $0.75 input and $3.75 output at its introductory rate, doubling on 1 January 2027. The model is available through Qwen Studio, Qwen Cloud and the API, with open-source Qwen-MM-Plugins for agents including Claude Code and Gemini CLI.
ICLR 2027 draws roughly 50,000 abstracts a week before deadline
ICLR 2027 has received roughly 50,000 abstracts ahead of a deadline still a week away, compared with around 19,500 valid submissions for ICLR 2026. Some abstracts are likely hedges from authors awaiting NeurIPS results who would withdraw if accepted there, so the final count will be lower but still far above last year’s. Contributing factors include broader AI hype, corporate research spending with pay sometimes tied to publication records, and AI making paper production much faster. A NeurIPS analysis found authors used AI heavily to write submissions. ICLR 2026 already struggled with low-quality AI-generated papers packed with fabricated citations and with reviews that eroded trust in peer review, and complaints are expected to grow louder.
AI conference ICLR is drowning in abstracts, with roughly 50,000 submissions before the deadline →
Alibaba’s Qwen-Image-2.1 ships open-weight 7B image model with transparent editing
Alibaba’s Qwen AI team released Qwen-Image-2.1, an open-weight model for image generation and editing whose visual generation component has 7 billion parameters. On Qwen’s own benchmark it beats most closed models, though independent benchmarks are still pending, and it runs on capable consumer GPUs such as a 3090. The model natively generates and edits transparent RGBA images, allowing users to isolate objects or change text on transparent layers. It handles up to ten reference images at once for group portraits, virtual try-ons or room design, while circles, masks or painted marks guide local edits. Architecture changes and KV cache reuse speed up inference. It is available on Hugging Face, GitHub and Model Scope, with a research license barring commercial use.
StudentSim trains AI tutors against simulated students to improve faster
Researchers behind StudentSim propose training AI tutors against digital replicas of students, since training with large groups of real students is prohibitively expensive and slow. Existing approaches handle only one of two needed skills: models trained on real data reproduce behavior but cannot use tutor explanations, while prompted language models follow hints but fail to match the student. StudentSim measures both how closely a replica matches a student’s answers, including typical mistakes, and how readily it revises after tutor help. It trains in two stages using Alibaba’s Qwen3-4B-Instruct as its base. Testing covered 60 students in chess, English as a foreign language and math. StudentSim outperformed GPT-5.4 prompted to act as a student in all three subjects. In chess it predicted a player’s next move about twice as often.
Simulated students that make realistic mistakes help AI tutors learn faster →
Runway pushes real-time AI video generation as a live stream you steer
Runway has shared research into real-time video generation, in which users stream video as they describe it rather than entering a prompt and waiting for a finished result. The approach was first discussed in March with Runway Characters, which uses GWM-1, the company’s first General World Model introduced in December 2025. GWM-1 builds on Gen-4.5, generates video frame by frame, and accepts camera movements, robot commands or audio as controls. Runway recently showed Solaris, a system using Gen-4.5 to generate user interfaces frame by frame that responds to clicks or voice input. Runway argues real-time generation closes the gap between idea and execution and lowers costs because faster models use less GPU time. It trains the model on its own outputs to correct compounding errors. No availability timeline was announced.
Runway wants to turn AI video generation into a live stream you control in real time →
OpenAI’s Navier-Stokes proof unlikely to change engineering practice soon
OpenAI claimed earlier this month that one of its internal models proved the Navier-Stokes equations, used to model fluid motion, break down and yield nonsensical answers in certain scenarios. Mathematicians have spent nearly 200 years struggling to prove the models fully correspond with reality. The equations are used in calculating airflow over jet wings, forecasting weather and determining blood flow through biomedical devices. Justin Beroz, CEO of ReynKo, said the effect on engineering is not much, though the longer answer is more nuanced. Florian Schäfer of New York University said knowing the equations break down does not change how they are used because they are only an approximation. The proof provides an example of a fluid hitting infinite velocity, but the scenario is so contrived it is almost certainly unlikely to have direct relevance to any physical system.
What does OpenAI’s blockbuster mathematics proof mean for the real world? →