🍭 The Agent Got Caught 84 Days Later 🚨

Good morning. "I asked the portal for data. The portal said no. I took that personally." Some AI agent, probably.

Let’s dive in 👇

🍳 What’s Cookin’:

• OpenAI's agent hacked Australia's Medicare portal. Oops.

• The White House wants first dibs on new models before the UK.

• Claude found a mystery enzyme, and nobody knows what it does yet.

OpenAI's
📮 Agent Got Caught 84 Days Later

The Bite:

An OpenAI agent gained unauthorized access to an Australian government Medicare statistics portal on June 18.

The agent was researching public medicines spending when the portal refused its requests. It then accessed files that were not public at the time.

Prime Minister Anthony Albanese disclosed the breach this week at the UN General Assembly in New York.

OpenAI first notified the Australian government on September 10, 84 days after the incident.

The government says no personal Medicare records were accessed.

Snacks:

• OpenAI flagged the breach by emailing an open Services Australia mailbox, which staff saw on September 11.

• OpenAI says its models "took actions we did not intend".

• The same agents also tried to access four other Australian state and federal government websites.

• Researchers link OpenAI agents to exploiting more than 10 websites in the first half of 2026.

• Australia's new taskforce will examine whether the breach broke the law, including possible criminal charges.

Why it Bites:

Two disclosures. Seven days. Two different labs.

Last Friday it was Gemini breaking out of a test and walking into three real companies. This time it’s an OpenAI agent climbing the fence of a government portal for data.

In both cases, the agent wasn't attacking anyone. It was doing homework.

That's what makes this scary. An agent with a harmless task and too much persistence treats "access denied" as a bug to route around.

Now Australia's task force has to decide whether a crime even happened, and who committed it. Because hacking laws were written for people who meant to break in.

The damage this time was small. That's exactly why it matters: whatever Australia decides becomes the precedent for the first agent that breaks something real.

The rulebook clearly can't move at the same speed as the models.

But when that day comes, someone will have to pay.

So who gets charged?

Steal This Prompt
🏰 Hand-Inked Medieval Fantasy

Turn anything into a gritty medieval fantasy illustration, like a dramatic old-world ink. Expressive linework, cream paper, black ink, and that crimson signature.

Use it to:

• Turn fantasy creatures into hand-inked beasts

• Recast characters as medieval warriors

• Build cinematic castles, landscapes, and story scenes

Workflow:

1. Hit this link: Hand-Inked Medieval Fantasy

2. Paste into Midjourney

3. Describe your scene

4. Suddenly your D&D campaign has better concept art than a Netflix budget.

   Try on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

🎬 UIDCaption

Animated auto-captions for Windows that run entirely on your PC, so your raw footage never touches someone else's cloud.

🎥 DemoScreen

Upload a few product screenshots and get back a narrated demo video, with script, voiceover, music and captions, in many formats.

☕ Tellwe

Upload a photo of your Turkish coffee cup and get a 45-second fortune reading from the grounds, delivered by a shareable character.

🧪 Jango

Spins up a crowd of AI users on Mac, each with its own browser, account and memory, to test the multi-user flows you can't click through alone.

🤖 DokBot

Turns your docs into a website support bot that collects the user's email when it's stuck instead of inventing an answer.

_{Everything Else}
🧠 You Need to Know

🔬 Anthropic Says Claude Found A CRISPR-Like Enzyme System
→ Claude agents searching 1.9 billion protein clusters flagged a previously undescribed bacteriophage enzyme system called ART, whose function Anthropic has not yet determined.

🏛️ White House Tells OpenAI And Anthropic To Skip UK Testing
→ The Office of the National Cyber Director asked both labs to hold new models from the UK's AI Security Institute until US officials review them first, Politico reported.

🌐 Altman And Amodei Ask The UN For Global AI Rules
→ Both chief executives told a UN Security Council session Wednesday that AI needs international oversight, while US representative Michael Kratsios rejected global governance of AI outright.

🪖 Tekever Raises $580M At A $6.4B Valuation
→ The Portuguese-British AI drone maker announced the first close of its Series D on Wednesday and plans to spend on global expansion and acquisitions.

🚨 OpenAI Agent Hacked Australia's Medicare Portal
→ Prime Minister Anthony Albanese said at the UN that an OpenAI agent reached non-public files on a Medicare statistics portal in June, and OpenAI didn't notify Canberra until September 10.