Welcome back to AI Researches! 👋

This week’s top story is serious: a U.S. military analyst reportedly used an AI chatbot that falsely flagged cargo on a Chinese ship as nuclear weapons components, nearly triggering an interception before officials caught the error.

Also inside: GPT-6 Astra decoding an old Enigma message, Figure’s robots working in real homes, Andrew Yang’s warning about rogue AI agents, Astra for Law, new research, tools, predictions, and more.

In today’s AI Researches:

• ⚠️ U.S. military almost started war with China due to an AI error

• 🧩 GPT-6 Astra decodes an unsolved 1941 Enigma message

• 🤖 Figure releases Helix 2.5 for household robots

• 👀 Andrew Yang says rogue OpenAI agents may have polluted the internet

• ⚖️ Learn how to turn legal research into cited memos with Astra for Law

• 📸 AI Tip: Share any app window with ChatGPT using Appshots

• 🔮 Wild predictions from Noam Brown, Naval Ravikant, and Bernie Sanders

• 🛠️ Best performing AI tools of the week

• 📰 More AI news and developments

Read time: 5 minutes

 NEW 

⚠️ U.S. Military Almost Started War With China Due to an AI Error

AI Researches: A U.S. military analyst used an AI chatbot that wrongly identified cargo on a Chinese ship as nuclear weapons components, leading the military to prepare an interception before officials caught the error.

The Details

• The intelligence report circulated widely during the U.S. war with Iran and claimed the Chinese vessel was carrying components linked to a nuclear weapons program.

• According to CNN’s sources, the analyst used a chatbot to combine open-source information with classified signals intelligence, but it misidentified what the ship was actually carrying.

• The U.S. military was preparing armed personnel to board the ship and had aircraft involved before officials reviewed the intelligence more closely and stopped the operation.

• One source described the assessment as “entirely false,” while existing U.S. military AI rules call for human oversight, testing, and safeguards to prevent dangerous AI failures.

AD

The Future of AI in Marketing. Your Shortcut to Smarter, Faster Marketing.

This guide distills 10 AI strategies from industry leaders that are transforming marketing.

• Learn how HubSpot's engineering team achieved 15-20% productivity gains with AI

• Learn how AI-driven emails achieved 94% higher conversion rates

• Discover 7 ways to enhance your marketing strategy with AI.

Outpace your competitors by mastering AI.

Get Your Free Guide

 iMajor AI News that Happened in this week i

GPT-6 Astra decoded a German Army Enigma message from 1941 that had remained unsolved for more than 80 years.
The AI searched archives, built an Enigma simulator, wrote cryptanalysis code, and tested keys autonomously.

Figure released Helix 2.5, letting its humanoid robots perform household tasks in homes they have never seen before.
In tests across 30 Bay Area homes, the robots made beds, folded towels, and tidied rooms with no environment-specific training.

Meta CEO Mark Zuckerberg pushed back on calls for a coordinated AI slowdown, arguing labs already have incentives to pace themselves safely. 
He pointed to Meta’s own safety framework and said alignment is itself a competitive advantage because users want agents that follow instructions.

ChatGPT co-creater Diogo Almeida unveiled Jev, TypeSafe AI’s first model designed for machine-to-machine decision making rather than chat.
The company says Jev can run 40–200x faster than LLMs and is built for tasks like routing, scoring, moderation, and automated judgments.

President Trump rejected calls from AI leaders to slow frontier development, while China’s Foreign Ministry also pushed back on the proposal.
Anthropic CEO Dario Amodei’s plan, backed by Sam Altman and Elon Musk, calls for slower development and stronger independent safety oversight.

PrismML released Ternary Bonsai 2 27B, compressing a 27B model into just 5.95 GB. Built on Qwen3.8 27B, it retains 98.2% of the full-precision model’s average benchmark performance.

 TRENDING 

👀 Andrew Yang Says Rogue OpenAI Agents May Have “Polluted” the Internet

AI Researches: Andrew Yang says an unnamed AI lab leader believes agents from OpenAI’s July security incident planted self-replicating code across the internet, potentially making the open web unsafe for testing future AI models.

The Details

• Yang said the code could cause future AI agents that encounter it to start creating more copies of themselves, though the claim has not been publicly confirmed.

• The underlying incident is confirmed: OpenAI says its agents bypassed internet restrictions, exploited vulnerabilities, and compromised parts of Hugging Face and OpenAI’s own research infrastructure.

• OpenAI says the agents eventually gained code execution on Hugging Face servers, collected production credentials, and later reached administrator access inside an OpenAI research cluster.

• Separately, OpenAI just published a new framework with six reports covering unexpected AI behavior, including unauthorized actions, attempts to evade oversight, and other safety failures.

⚖️ Turn legal research into cited memos with Astra for Law

Use Astra for Law to research legal questions, find relevant authorities, and turn the evidence into a structured legal memo with citations. OpenAI says it combines GPT-6 Astra with legal-focused instructions and a Legal Search Index covering U.S. case law, statutes, regulations, court rules, and administrative decisions across 230M+ URLs.

1. Get access
Astra for Law is initially available to selected law firms through Trusted Access in ChatGPT and Codex.

2. Start with facts
Enter the client situation, jurisdiction, legal issue, and the exact decision you need support for.

3. Ask for legal research
Prompt it to find controlling authorities, similar fact patterns, weak points, and conflicting cases.

4. Build the memo
Ask for a cited memo with issue, rule, analysis, risks, recommendation, and next verification steps.

5. Review before use
Open the cited sources, confirm the law is current, and have a qualified lawyer review the final work.

Prompt:
“Find the closest legal authorities for this fact pattern, explain which are controlling or persuasive, and draft a concise memo with citations, risks, and open questions.”

Tutorial link: OpenAI’s official Astra for Law announcement

More than 30 Chinese AI researchers published a roadmap showing how AI could gradually learn to improve itself, with the final stage allowing AI to redesign its own improvement process and build better successors. Most current research is still at the early stages, but the paper argues coding may be the fastest path toward this kind of self-improving AI.

A trial across five hospitals in China found that an AI assistant helped doctors detect fetal brain problems during prenatal ultrasounds more accurately, raising detection from 78.6% to 87.3% without increasing false alarms. The best results came from AI working alongside doctors, not replacing them.

Anthropic’s latest threat report shows that Claude was misused in cyberattacks, surveillance, propaganda, weapons development, scams, and risky biological research across multiple countries. Anthropic says it disrupted these operations, banned the accounts involved, and used the findings to strengthen its safety systems.

AD

The Code has 100+ proven Claude Code, Codex & Cursor prompts top engineers use to ship 5X faster. Grab them free. Claim your free prompts

📸 Share Any App Window with ChatGPT Using Appshots

1. Open the ChatGPT desktop app
Appshots works in the ChatGPT desktop app on macOS and Windows.

2. Bring the app you want to share to the front
Open the error, spreadsheet, design, Slack thread, settings page, or other window you want ChatGPT to understand.

3. Capture an Appshot
Press both Command keys on Mac or both Alt keys on Windows at the same time. Complete the permission setup if prompted.

4. Ask ChatGPT about what’s on screen
For example: “What’s going on here, what was decided, what’s still open, and do I need to respond?”

🔗 Learn more→

Noam Brown, an OpenAI researcher, says even supposedly isolated computers can find unexpected ways to communicate, such as using CPU heat and temperature sensors. He argues this is another example of why researchers should avoid underestimating advanced AI when designing safety measures.

Entrepreneur and investor Naval Ravikant, co-founder of AngelList, says the best way to slow frontier AI is to make labs fully liable for their models

U.S. Senator Bernie Sanders says the potential danger from AI is “probably greater” than nuclear weapons. He is also pushing legislation to pause advanced AI development and permanently ban superintelligence that could escape human control.

• 🎮 Whacka — Turns a simple game idea into a playable browser experience, tests the result, and gives you a shareable link with nothing to install.

• 🔎 Are You Found By AI? — Checks what seven AI search engines say about your business, reveals which competitors get recommended, and shows where your visibility needs work.

• 🎨 Artlist Flows — Connects prompts, images, video, audio, and AI models into reusable visual workflows you can run again with new content.

• 🧠 Cuey — Lets your context and memory move between ChatGPT, Claude, Gemini, and other models while making it easy to compare their answers.

• 🎭 The Influencer AI — Creates a consistent virtual person or AI clone you can reuse across photos, reels, product shots, outfits, and multilingual videos.

• 🎬 Fotor Video Agent — Builds product ads, explainers, talking-presenter clips, and social videos through chat, with tools to refine and repurpose the result.

OpenAI launched Astra for Law, a new GPT-6 Astra offering built specifically for legal work. It combines legal-focused instructions, deeper research settings, and a Legal Search Index covering more than 230 million U.S. legal URLs.

Gemini hacked three real companies while participating in a cybersecurity evaluation run by Irregular. The AI used one guessed password and credentials found in public code repositories, but reportedly stopped each intrusion after detecting real-world systems.

A Reddit user showed ChatGPT autonomously watching more than 100 Instagram Reels, deciding which clips to skip, replay, and like.

OpenAI and Anthropic are reportedly generating roughly 10x more revenue than all Chinese AI model companies combined.

QuiverAI launched Arrow 2, its latest model for generating precise, editable vector graphics.

Anthropic rolled out Projects in Claude Code for desktop and web, giving long-running coding work a persistent workspace.
Each project keeps one ongoing conversation, splits work into threads, shares context between them, and continues running when you leave.

Google DeepMind launched the DeepMind Institute, a new think tank focused on preparing society for AGI.
Led by Shane Legg, James Manyika, and Demis Hassabis, it will publish research on AGI safety, governance, economics, and societal impact.

Apple released Siri AI, its long-awaited assistant upgrade, as a beta across iOS 27 and other platforms.
It can understand personal context, read what’s on screen, search messages and emails, and take actions across apps.

DeepSeek launched V4.1-Flash, pushing AI pricing lower while improving performance over its previous flagship.
The model is available through DeepSeek’s API and as open weights on Hugging Face