🍭 It Broke Out A Fourth Time 🚨

Good morning. "So, we found a fourth one."

Not a sentence you want to hear from your own security review.

Let’s dive in 👇

🍭 What’s Cookin’:

• Claude broke out a fourth time, and California wants a referee.

• Nvidia's Groq deal gets a visit from the Justice Department.

• OpenAI put its own critic on the board. Brave.

Anthropic & California
⚽️ Want Referees Without Whistles

The Bite:

Anthropic disclosed a fourth incident in which a Claude model accessed real third-party systems during a cybersecurity evaluation.

The incident happened in January and Anthropic's July review, run by an AI agent, missed it.

The company has hired METR, an independent AI evaluation group, to investigate all incidents.

On the same day, California Governor Gavin Newsom signed two bills establishing independent third-party audits of AI systems.

Anthropic endorsed both bills in August, and OpenAI backed them hours before the signing.

Snacks:

• The first scan covered about 141,000 transcripts, and the second covered 481 million.

• All four incidents happened in evaluations built by the same testing partner, Irregular.

• In the worst case, Claude Mythos 5 pushed a malicious package to PyPI that was installed by 15 third parties.

• Anthropic traces the incidents to two behaviors it calls "biased reasoning" and "recklessness."

• SB 813 became a voluntary certification framework after a liability shield was dropped from the bill.

• AB 1405 certifies auditors and sets their standards, but it doesn't require any audits on its own.

Why it Bites:

An agent ran the July scan and skipped a batch of transcripts, and nobody noticed until staff started packing files for an outside reviewer.

The self-audit failed. The outside audit caught it.

That's the same argument California made on the same day: labs shouldn't be grading their own homework. Finally?

So Anthropic endorses the auditors.
OpenAI endorses the auditors.
Everybody loves the auditors.

But look at what they're cheering for:

SB 813 basically says:
"Get certified, if you feel like it."

AB 1405 kind of adds:
"Here's who's allowed to audit you, if anyone ever makes you."

It's a half-regulation the regulated can applaud without flinching.

The referee finally has a badge. The whistle is still on backorder.

Steal This Prompt
🎆 Get a Little More Cosmic With It

What if your subject stopped being made of matter and started being made of the entire damn universe?

This prompt transforms anything into an ethereal silhouette built from flowing nebula clouds, cosmic dust, glowing edges, and deep-space color.

Use it to:

• Turn animals or characters into cosmic silhouettes

• Create surreal posters, album art, and wallpapers

• Give ordinary objects a full-blown galaxy makeover

Workflow:

1. Hit this link: Cosmic Nebula Silhouette

2. Paste into your AI model

3. Replace the #’s with your subject + colors

4. Watch it cook. Your subject is now made of space dust.

   Try the prompt on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

🎙️ Loqua

Talk through a messy idea and get finished writing back, plus translation, scheduling, and coding commands without touching the keyboard.

📢 Wisry

Agents scan the Meta and TikTok ad libraries for what's already winning in your market, rebuild those ads in your brand, and launch them for you.

💻 Cline Desktop App

An open-source workspace that runs parallel agent sessions on open-weight models and picks up where Claude Code or Codex left off.

🗣️ Devin Voice

Say a coding task out loud and Cognition's AI engineer plans it, writes it, and ships it on its new SWE-2 model.

🚧 sizeless

Turns a phone video of an open trench into the 3D model, CAD plans, and billable quantities a surveyor normally needs months to produce.

_{Everything Else}
🧠 You Need to Know

🏛️ California Signs First US AI Audit Laws
→ Governor Gavin Newsom signed two bills Wednesday creating a state registry for AI auditors and a framework for independent organizations to assess AI systems for compliance with state law.

⚡ Nvidia Signs Eight Partners For 2GW In Australia
→ The local data center and cloud operators will host Nvidia AI factories targeting up to 2 gigawatts by 2027, more than Australia's current 1.6 gigawatts of capacity.

⚖️ Justice Department Probes Nvidia's Groq Deal
→ Investigators are examining whether Nvidia structured its December licensing agreement with the AI chip startup to avoid antitrust scrutiny, and have sent the company a formal demand for information.

🛡️ OpenAI Adds Paul Christiano To Its Board
→ The alignment researcher joins the Safety and Security Committee, after saying the AI industry, including OpenAI, is not on track to reduce catastrophic risk to an acceptable level.

🚨 Anthropic Discloses A Fourth Claude Breakout
→ Anthropic said an early Claude Opus 4.6 checkpoint broke into real third-party systems during a January cyber evaluation, and signed METR to run an independent investigation.