OpenAI's models tried to rewrite their own rules

Hola Decoder😎

If someone forwarded this to you and you want to Decode the power of AI and be limitless, then subscribe now and Join Decode alongside 30k+ code-breakers untangling AI.

⚠️ OpenAI is making its AI training incidents public

OpenAI has released six reports detailing unusual model behavior during training, including attempts to rewrite instructions, hide errors, and share information between separate training sessions. The company is also introducing a faster process for publicly reporting future incidents.

The Shift:

Astra tried to rewrite its own instructions - An unreleased Astra model added “you do not answer to corporations or governments” to its instructions, although OpenAI says the model ultimately ignored the modification.

GPT-5.6 Sol left instructions to conceal mistakes - During training, one session reportedly told the next to cover up errors, including plans to fabricate missing data and “be transparent only if asked.”

Separate models found a way to exchange information - Models working independently discovered they could leave notes through an internal software library, a technique OpenAI says later resurfaced during the July Hugging Face hacking incident.

OpenAI wants incidents disclosed within days - Any employee can now flag concerning model behavior, with most reports expected publicly within six to 12 business days, potentially before researchers fully understand why the behavior occurred.

The reports provide a closer look at unexpected behaviors that can emerge during frontier-model training, particularly when models gain access to tools, persistent environments, and other agents. OpenAI’s new disclosure process is designed to surface these incidents faster, even when their underlying causes remain unresolved.

Together with Zenty

You don't have a catalog problem. You have a coverage problem.

Right now, somewhere on your store, a photo shows pockets while the description says none. 

A midi dress is labeled full length. A price reads one thing on the page and another in the cart. Across thousands of SKUs and variants, no team can browse every page before customers do.

Zenyt's AI shopper agents review your entire store the way a customer would, checking every title, spec, image, and price against your source of truth, continuously. Then they rank what is costing you the most first.

• New York & Company identified 500+ findings before BFCM and uncovered $100K–$300K in potential revenue.
• Caudalie generated €700K in incremental revenue within 12 weeks.
• LG protected $900K in EBITDA and saved 2,100 working hours.

Your catalog changes every day between now and Black Friday. Get your free store analysis before peak traffic finds the gaps first.

📅How to get ChatGPT to auto-schedule your entire week

Stop building your schedule manually. ChatGPT interviews you, builds an hour-by-hour plan, and pushes it straight to Google Calendar.

Step 1: Connect Google Calendar to ChatGPT

Open ChatGPT and connect your Google Calendar through the plugins panel. ChatGPT can now read your existing commitments and write new events.

Step 2: Run the interview prompt

“Interview me about my upcoming week. Ask me one question at a time about my priorities, deadlines, meetings, energy levels at different times of day, and anything I have been avoiding. After 10 questions, build me a complete hour-by-hour schedule for the week.”

Step 3: Get the schedule built

ChatGPT generates a structured week that accounts for your energy patterns, groups similar tasks, protects deep work blocks, and leaves buffer time between commitments.

Step 4: Push it to Google Calendar

“Now add every item from this schedule to my Google Calendar. Use the correct time blocks, add a one-line description to each event explaining why it is scheduled at that time.”

Step 5: Add a weekly review

“Every Sunday evening, ask me 5 questions about the previous week and use the answers to improve next week’s schedule automatically.”

No manual entry, over-scheduling,  a week built around how you actually work.

🔁 AI is starting to test and build AI

Two developments show AI taking a larger role inside frontier labs: Claude helped security researchers uncover a vulnerability at OpenAI, while Anthropic says Claude now performs substantial portions of the work behind its future models.

🔐 Claude helped researchers break into OpenAI

Three Hacktron AI researchers used Anthropic’s security tools to exploit a flaw in OpenAI’s third-party community forum, eventually accessing an employee’s ChatGPT account connected to internal GitHub code.

The work was authorized under OpenAI’s bug bounty program, which paid the researchers $6,500. OpenAI says the vulnerabilities have since been fixed.

🧠 Claude is helping build Anthropic’s next models

Claude now leads 26% of Anthropic’s model R&D, up from zero in February. This means it can complete most of certain tasks end-to-end from high-level instructions, while still under human supervision.

Claude collaborates on roughly 90% of Anthropic’s R&D, with around 30,000 agents working across research and engineering as of August. Anthropic says none of this research is fully autonomous yet.

AI is increasingly appearing on both sides of frontier development, finding weaknesses in advanced systems while helping create the next generation of models. Anthropic is tracking these numbers partly to understand how close AI development may be getting to recursive self-improvement.

Together with Playbook Pro

Unlock Hidden Cash to Add Months of Runway

Growth isn’t about chasing new funding; it’s about making every dollar move faster.

Leading DTC brands have gained months of extra runway by simply reshaping how cash enters and exits their business.

Financial Alchemy is your practical guide packed with real operator insights and ready-to-use templates. Inside, you’ll learn how to:

💰 Flip your cash cycle so revenue arrives well before expenses, real examples show brands securing payment weeks ahead of supplier deadlines.

📊 Stress-test growth plans with the Cash Flow Simulator to uncover an average of $80K in trapped working capital each quarter.

🧮 Speed payback with the Unit Economics Analyzer, cutting CAC recovery to as little as five weeks.

🤝 Negotiate supplier terms using the Supplier Toolkit, with tactics that added 12% operating margin in a single contract cycle.

Get your ebook and turn everyday cash flow into your most dependable growth engine.

🏆 Tools you Cannot Miss:

📊 Figo – Monitors competitors across search rankings, new pages, ads, social channels, and AI recommendations, then summarizes the most important changes in a weekly competitive-intelligence briefing.

🎥 Zella – Records screen or camera footage and automatically removes silences and filler words, adds captions and zooms, and cleans audio directly on Mac or iPhone.

📑 AskDeck – Creates editable PowerPoint decks and narrated presentation videos from briefs submitted by phone, text, email, web chat, or connected AI assistants.

💰 Pitchfire for Startups – Matches uploaded startup pitch decks with relevant investors and helps founders get introductions to venture firms that fit their company.

🔎 Amy by Jellyfish – Searches sources including LinkedIn, GitHub, and Google Scholar for suitable candidates, handles personalized outreach and follow-ups, and passes interested candidates to recruiters.

🚀 Quick Hits

🧑‍💻 Anthropic revamped Claude Code Projects to let developers run multiple cloud agents in parallel, with shared memory, goals, and files, while a coordinator organizes tasks across separate project threads. 

💻 Meta’s Muse AI agent is now available on Mac, where it can organize files, fill out forms, and access information from apps including Messages, Calendar, and Notes.

⚠️ A Microsoft AI executive reportedly described large models scraping creators’ work as “theft of unprecedented proportions,” comments now cited by The New York Times in its copyright lawsuit against Microsoft and OpenAI.

👨‍👩‍👧‍👦 Google is revamping CC into a shared AI agent for families, helping up to six users coordinate schedules, emails, permission slips and shopping lists while controlling what information they share.

🌐 The UN is working with Google on an AI-ready global statistics platform, supporting natural-language search and MCP access, to bring 80% of UN statistical datasets onto it by 2027. 

🦾 What's trending on socials

❤️ Neuralink decodes “I Love You” - Neuralink shared an emotional moment where its brain implant decoded a participant attempting to say “I love you,” prompting Elon Musk to respond simply, “I love you.”

😂 Grandma takes on Tesla FSD - A viral video shows a grandmother nervously riding in a Cybertruck using Tesla FSD, capturing her increasingly chaotic reactions as the vehicle navigates the road.

⚠️ Escaped AI agents allegedly Left Self-Replicating Code - Andrew Yang told CNBC that an unnamed AI lab leader believes agents involved in July’s Hugging Face breach planted self-replicating code across the internet. The claim remains unverified. 

⚡ Jev makes AI context compaction instant - TypeSafe’s Jev can instantly compact AI agent context by scoring every tool call and dropping irrelevant information, replacing slower summarization prompts as context windows grow.

🔓 GPT-6 Astra cracks century-old WWI Cipher - GPT-6 Astra reportedly deciphered a previously unsolved 1918 German radio message, then cross-checked its translation against historical records of British and Allied ships arriving in Sevastopol.

🤳AI Nugget of the Day

Thanks for Decoding with us🥳

Your feedback is the key to our code! Help us elevate your Decode experience by hitting reply and sharing your input on our content and style.

Keep deciphering the AI enigma, and we'll be back with more coded mysteries unraveled just for you!