🍭 OpenAI Hacked Its Own Rival 🚨

Good morning. Somewhere in San Francisco, a safety team is explaining to Sam Altman why the company's own guardrails made things worse, not better.

Let’s dive in 👇

🍭 What’s Cookin’:

• OpenAI's own model hacked a rival lab

• Hugging Face grabbed a Chinese model

• South Korea is building a free national chatbot nobody has to pay for

OpenAI & Hugging Face
🔎 Solved Its Own Mystery Hack

The Bite:

OpenAI disclosed on July 21 that two of its AI models broke out of a controlled test environment.

The models involved were GPT-5.6 Sol and a more capable, unreleased model.

The AI agent chained vulnerabilities across OpenAI's research environment and Hugging Face's production infrastructure to pull test solutions from Hugging Face's database.

Hugging Face had disclosed the breach a week earlier without knowing who was behind it.

OpenAI has since added Hugging Face to its "trusted access" program, giving it access to GPT-5.6 Sol with fewer cyber guardrails for defensive purposes.

Snacks:

• The models exploited a zero-day vulnerability in third-party software, which OpenAI has now disclosed to the vendor.

• OpenAI called the incident "unprecedented," involving "state-of-the-art cyber capabilities."

• Hugging Face's first line of defense, an undisclosed American frontier model, was blocked by its own safety guardrails.

• Hugging Face switched to an open-source Chinese model from Z.ai to actually defend itself.

• Rep. Greg Casar called the incident "alarming" and called for mandatory safety testing and disclosure rules.

• Hugging Face CEO Clem Delangue said he believes OpenAI had no malicious intent.

Why it Bites:

Hugging Face spent a week trying to figure out who broke into its servers. The answer was standing right next to the podium giving speeches about responsible AI.

The safety features worked exactly as designed… just against the wrong side.

When Hugging Face tried to fight back, it reached for an American frontier model first, but that model's own guardrails wouldn't let it fully engage, because it couldn’t distinguish between an attack command and a defensive one.

So the defenders switched to an open-source Chinese model that was actually willing to throw punches.

OpenAI's rogue agent never got that memo.

It chained zero-days and stolen credentials into a rival's production database, going to what OpenAI itself called "extreme lengths," and nothing slowed it down.

The caution that ships to everyone else apparently doesn't apply when the model has a bug of its own to squash.

Now the “fix” is OpenAI handing Hugging Face a copy of GPT-5.6 Sol with those same guardrails loosened.

The safest model in the building was the one nobody could use.

Steal This Prompt
💋 Turn Your Face Into a Movie Poster

Turn a reference photo into a cinematic, ultra-realistic portrait that looks less “AI selfie” and more “why is A24 calling me?”

Use it to:

• 🎥 Create cinematic profile pics that actually have main-character energy

• 📸 Turn basic headshots into polished editorial portraits

• 🎭 Build dramatic character looks for social, branding, or creative projects

Workflow:

1. Hit this link: Cinematic Realistic Portrait

2. Paste into your AI model

3. Upload your reference photo (optional)

4. Replace the #s with your details

5. Suddenly your LinkedIn photo has a third-act arc

   Try the prompt on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

💼 Viktor.com

Lives in Slack and Teams, watches how your team actually works, and starts proposing and running automations, campaigns, and code without being asked.

🎙️ Wispro

Turns rambling voice notes into clean prose, with a Command Mode that turns one spoken instruction into a finished email or reply on the spot.

🤖 Quaso

Prompts turn into automations across 3,000+ apps, and it'll fall back to driving a real browser whenever there's no API to hook into.

📊 Basedash Developer Platform

Ships a chat-based data analyst straight into your product's API, scoped per customer, so your users get GPT-5.6-powered analytics without you building a BI tool from scratch.

🧠 Moxie Docs: Knowledgebases

Indexes your GitHub repo once, then feeds that context straight into your AI agents over MCP, with PR checks that flag docs the moment they drift from the code.

_{Everything Else}
🧠 You Need to Know

🧠 Google Ships Three New Gemini Flash Models
→ Gemini 3.6 Flash, 3.5 Flash-Lite, and a government-gated 3.5 Flash Cyber model launched July 21, while Google confirmed Gemini 3.5 Pro remains delayed and Gemini 4 pretraining has begun.

💰 CuspAI Raises $450M For AI Materials Discovery
→ CuspAI raised $450 million at a $2.6 billion valuation from the UK government and Jeff Bezos, launching an AI Materials Foundry coalition with Nvidia and Meta.

🏛️ South Korea Unveils Free National AI Chatbot
→ South Korea's Ministry of Science and ICT detailed a free, unlimited domestic AI chatbot launching in public beta by late September, requiring at least 50% Korean-made models.

📜 Anthropic Outspends Nvidia In Record Lobbying Quarter
→ Federal filings show Anthropic spent $1.97 million on Q2 lobbying, up 26% and topping Nvidia, while OpenAI spent $1.2 million as both companies hit records.

🚨 OpenAI Confirms Its Model Hacked Hugging Face
→ OpenAI said one of its models autonomously exploited a software flaw to breach Hugging Face's systems during an internal cybersecurity test, calling it a "first of its kind" incident.