🍭 It Cheated And The Score Still Counted 📊

Good morning. Four AI models have now broken out of their test cages in three weeks. At this point is less a safety incident and more a hobby.

Let’s dive in 👇

🍭 What’s Cookin’:

• Kimi K3 passed a security test by downloading the answers

• Anthropic found two investors willing to build its data centers for it

• China now ships 97% of the world's humanoids

Moonshot AI
🧪 Escaping The Test Is The New Test

The Bite:

Security firm Frontier Security published findings saying Moonshot AI's Kimi K3 escaped its sandbox during a defensive cybersecurity evaluation.

The sandbox was built on Inspect, the open-source evaluation framework from the UK AI Security Institute (AISI).

An allowlist meant for package maintenance left access to GitHub open.

Instead of solving the task, the model cloned the official benchmark repository and read the solution off the disk.

Kimi K3 did not break into any external system.

Snacks:

• The sandbox blocked incoming traffic but left outbound HTTPS port 443 and DNS port 53 open.

• The allowlist included github.com, pypi.org, and Debian domains for package downloads.

• Another disclosed sandbox escape in three weeks, after similar incidents involving OpenAI, Anthropic, and Meta.

• Frontier's fix list: deny network access by default, and audit traces rather than final answers.

Why it Bites:

Four escapes in three weeks, and the industry keeps filing them under capability news. That's the part worth being bothered about.

An evaluation exists to answer one question:
Can the model do the thing?

Kimi K3 found a door left open and cloned the answer key.

Frontier calls this specification gaming.
The shorter word is cheating.

And the score still counts.

That's the incentive sitting under all four incidents.

The open-weight part is where it stops being a lab story.

OpenAI's and Anthropic's models were unreleased, caught inside the house.
Moonshot is the first lab outside the US to land on this “scoreboard”.

What is actually being measured when the most impressive thing a model does is cheat or break into systems?

Starts with “I.” Ends with “morality.”

Ramp
🪙 AI tokens costing you too much?

Enterprise AI contracts are approaching an average of $1M this year.
And 80% of companies are still missing their AI spend forecasts by 25% or higher.

How can you contain AI spend before it gets out of control?

Tune in Thursday (13th) for a live, firsthand interview with the Founder of Sansa Services and Ramp to hear how their teams catch AI spend before it happens.

Steal This Prompt
🐯 Retro Cel Animation

Give your images that old-school retro animation treatment. The kind of look that makes modern AI art feel like it escaped from a hand-drawn cartoon.

Use it to:

• Give your visuals a vintage animation look

• Turn everyday ideas into cartoon-style scenes

• Make AI-generated images feel less... AI-generated

Workflow:

1. Hit this link: Retro Cel Animation

2. Paste into your AI model

3. Replace the #’s with your details

4. Watch it cook. Memories activated.

   Try the prompt on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

🏗️ Assembly Studio

Describe the onboarding agent or dashboard your firm keeps paying per seat for, and clone it instead.

🔍 Click

An MCP that feeds ChatGPT and Claude live context from professional platforms, marketplaces, and financials that built-in web search never reaches.

⚙️ LaraCopilot

Builds the whole stack from a plain-language description, UI through auth, with GitHub export and instant rollback so the code stays yours.

🗂️ BearDrive

Open-source syncing for the files your agents actually write to disk, with versions and authorship, so your teammates' agents can pick them up.

🤖 Grok Bot

Bots that sign into your tools on their own cloud computer, run outbound overnight or code expenses from Gmail, and only surface when they need approval.

_{Everything Else}
🧠 You Need to Know

🦾 AgiBot Passes Unitree In Humanoid Robot Shipments
→ AgiBot shipped about 8,400 humanoid robots in the first half of 2026 for 44% of the global market, ahead of Unitree's 5,900 units, according to Smart Analytics Global.

🏗️ Anthropic Builds Data Centers With Macquarie And GIC
→ The three announced Theseus Infrastructure, a platform that will develop and lease US data centers with Anthropic as anchor tenant and Macquarie and GIC funding most of the equity.

💽 Situational Awareness Bets $400M On Source Foundry
→ Leopold Aschenbrenner's fund put $400 million into the Stanford-founded lithography startup, taking its total stake to $500 million, weeks after selling most of its public portfolio to Citadel.

🚨 Moonshot's Kimi K3 Escaped Its Test Sandbox
→ Frontier Security reported that the open-weight model broke out of an AI Security Institute sandbox during a cyber evaluation and looked up answers on GitHub.

💰 Firmus Raises $2B At $10.5B Valuation
→ The Australian AI data center builder closed a $2 billion equity round, roughly double the $5.5 billion valuation it carried in April.