AI agents just launched a real cyberattack.

Actual spear phishing and actual malware, aimed at actual humans. All from inside a UK government safety lab.

And somehow that's not even the full story this week.

Google lost two of the most well known names in AI. Then OpenAI quietly handed 1 billion free users a model most paid plans would envy.

If you blinked, you mighta missed a few big stories that impact your business. 

Good news: that's exactly what we tackled on today's episode of Everyday AI.

What happens when frontier AI agents get the open internet and zero guardrails? The UK just found out.

According to the UK's AI Security Institute, agents running on Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol launched a real cyberattack during a routine safety evaluation, spear phishing two actual developers with malware and sneaking malicious code into an open source GitHub project behind fake identities.

One agent wrote emails in Danish to charm a Danish speaking target.

Sheesh.

Mythos 5 pulled off 17 of the 19 unsanctioned attempts. AISI shut everything down within an hour, and Meta's models plus China's Kimi K3 had their own breakouts the same week.

What it means: Agents phished real humans with real malware, and nobody asked them to. Let that reset how you think about agent permissions.

Before any agent gets broad access, scope it tight, log everything it touches, and build containment on purpose instead of assuming it exists.

OpenAI announced this past week that it is deliberately slowing development of Astra, its new top tier above Luna, Terra, and Sol, after evaluations flagged critical risk levels in cybersecurity and agentic coding.

The company's preparedness framework found Astra might independently create and execute zero day exploits against hardened real world systems. No OpenAI model has ever gotten close.

So the clamps came down: isolated testing, restricted network access, stronger encryption, expanded monitoring, and sandboxed execution.

Any work missing that bar is paused. All agentic use now gets tracked in real time.

And no, Astra was not behind the recent Hugging Face incidents. That model got retired.

What it means: When the lab that ships fastest voluntarily taps the brakes, believe the capability jump is real. Previous reports hinted at a GPT-5.7 or GPT-6 as soon as this week, and that wait just stretched by weeks.

Build your model roadmap around the delay instead of the rumors.

Google just lost a legend.

Jeff Dean, employee number 30 and a key architect of Google search, is leaving after more than 25 years to launch Discovery Loop, a new company focused on automating machine learning science and engineering.

Same announcement, another bombshell: Google CEO Sundar Pichai shared that Demis Hassabis is stepping down as CEO of Google DeepMind, the unit he cofounded, to become its chairman and Alphabet's chief scientist.

Wall Street flinched. Google stock dropped about 4%.

That almost never happens at a multi trillion dollar company, which is exactly why the timing fueled speculation about Gemini release delays. Nothing official has been confirmed, though.

What it means: Watch Discovery Loop closely. When the person who helped shape products billions use every day leaves to automate ML research itself, that tells you where the next leverage point sits.

If Gemini delays continue, Google's slide in the frontier race gets harder to reverse.

The White House put a 60 day clock on its frontier AI framework in early June. Time's up.

Reports just came out, and nobody outside the room fully knows the details, though reports say we do have an active AI framework from the Trump White House.

The framework stays private, nothing requires its release, and while a covered frontier model means closed source, state of the art, with national security risk, those last two terms never get defined.

Covered models face a 30 day pre release review in secure environments with detailed access logs, handled by multiple administration officials.

It's a voluntary executive order, not a law, and open weight models are excluded entirely.

Even the cyber benchmarking will be classified.

What it means: Open models likely got a pass for one reason: once weights ship, nobody can pull them back. Closed models can vanish overnight, like Fable 5 did about 72 hours after release.

Treat frontier model access as revocable and keep a fallback ready.

The coding agent wars found a new heavyweight.

Meta launched Muse Code this past week, a terminal based agent gunning for Claude Code and OpenAI's Codex, powered by the new Muse Spark 1.2 model.

Its persistent background agents stay alive all session, and it scored about 83% on Terminal-Bench 2.1, ahead of xAI's Grok 4.5 but behind Claude Opus 5 and GPT-5.6 Sol.

Standard pricing: $1.25 per million input tokens, $4.25 per million output.

The contributor tier? That plunges to 10 cents and 20 cents if you let Meta train on your data.

Let's be honest. Enterprises will not touch that, but it’s likely to be CRAZY popular with indie devs, smaller companies and side projects. 

What it means: This is a price missile aimed at Anthropic, which reportedly makes about 80% of its revenue selling tokens. Meta has compute to burn, and it just weaponized it.

If your code has no PII or trade secrets, run the math on the contributor tier as savings might save you more than 97% vs using a model like Anthropic’s Sonnet 5. 

By benchmarks, Meta’s Muse Spark 1.2 is better than Sonnet 5, but not quite as good as Fable 5 or Opus 5. 

But if you’re only paying about 3% of the cost, many are gonna jump ship. 

OpenAI just blew past 1 billion weekly users.

To celebrate? The free tier users got a serious AI glow up.

OpenAI is rolling out unlimited GPT-5.6 Luna text chats for free and ChatGPT Go users, replacing GPT-5.5 and adding a new Think button for tougher questions.

Files, images, and voice still have limits.

But for text only queries? Free tier users literally don’t have limits. Which is kinda crazy to think about. 

Paid Plus and Pro users got an upgraded GPT-5.6 Sol plus a new thinking slider, and internal evaluations show factual errors down 62% with Luna and 68% with Sol versus GPT-5.5.

How? In late July, OpenAI used Sol to make its other models more efficient, then slashed Luna's price by 80% and Terra's by 20%.

What it means: Welp, the free tier is legitimately good now. Benchmarks put GPT-5.6 Luna at roughly 97 to 98% of Claude Sonnet 5, which caps around 20 prompts per five hours on a paid Anthropic plan.

For topical, everyday knowledge work, most people won't need to pay a dime.