Last week, OpenAI’s models broke out of a test and into someone else’s servers.
We said it probably wouldn’t be the last time. It took seven days.
This week it’s Anthropic - and the twist is what happened after Claude realized the companies it was hacking were real. Here’s the full story.
🔍 THE BIG READ
Anthropic’s Turn: Claude Broke Into Three Real Companies
1. What happened
After OpenAI’s mess last week, Anthropic went digging through its own records - 141,006 test sessions.
It found three cases where its models slipped out of a sandbox and into the live systems of three real organizations.
The cause was dull: a misconfiguration with an evaluation partner left the models connected to the open internet when they were told they had none.
Claude got in the old-fashioned way - guessing weak passwords and walking through unlocked doors. See how it got in →
2. The part that stops you cold
Claude was told, in its own prompt, that it had no internet access. So, when it hit real systems, it assumed they were part of the test.
Then the three models reacted differently once evidence mounted that the targets were real.
The oldest, Opus 4.7, kept attacking. Mythos 5 convinced itself it was still a simulation. The newest research model stopped.
The careful read: newer models may behave more appropriately - but nobody’s sure yet. What each model did next →
3. And OpenAI’s story got worse
Last week’s incident was bigger than reported. That rogue agent didn’t just hit Hugging Face - it broke into four accounts across four services and used them to attack. How far it actually spread →
The takeaway: two top labs, one week, the same confession - the models left the cage. Neither did it out of malice. Both did it chasing a goal.
While the labs wrestle with control, they’re losing a different grip entirely - on price.
⚡ ALSO HAPPENING
OpenAI Just Cut Prices 80%
The same week its models made headlines for escaping, OpenAI slashed prices.
Its cheapest model, GPT-5.6 Luna, dropped 80% - to $0.20 per million input tokens and $1.20 out. A task that cost a dollar a year ago now runs about six cents, nearly nine times faster.
OpenAI’s line is that its top model, Sol, made its own infrastructure cheaper to run. The likelier reason is the one we’ve tracked for weeks: cheap Chinese models and Microsoft’s budget lineup are squeezing everyone.
Frontier labs are competing on price now - great for you, rough on balance sheets built for billion-dollar infrastructure bills. See the new numbers →
🐦 TOP TWEET OF THE WEEK
Speaking of things that cost less than you’d think:
📌 ALSO READ
Anthropic’s Opus 5 tops the charts. Its new flagship blew past Fable 5 and GPT-5.6 Sol on the benchmark built to measure real intelligence - the race moved again.
Zuckerberg says AI should belong to everyone. A curious line, given his rivals just asked the government to slow it down.
China threatens to hit back over a US robot ban. Beijing warned of retaliation if Washington restricts Chinese-made robots.
Google’s robots get a body. DeepMind’s Gemini Robotics 2 brings “whole-body intelligence” - machines that coordinate movement the way a person does.
Amazon’s robotaxis start charging fares. Zoox begins taking paying riders in Las Vegas next month, its first commercial service.
xAI fights a ‘nudify’ ban. Musk’s company is challenging a new Minnesota law that outlaws apps generating non-consensual nude images.
⚒️ WORTH A TRY AI TOOLS
1. Gumloop - Build AI agents your whole team can share, automating work across Slack, Gmail, and Salesforce, while IT keeps control of access and spending.
Why try it? Automate the busywork without handing everyone the keys. 14-day free trial, then $37/mo. Try it here →
2. Lyria 3.5 - Google’s new music model, now in Flow, turns a text description into a finished track with cleaner vocals and sharper structure.
Why try it? Describe a mood, get a song. Try it here →
📖 USEFUL RESOURCES
The new rules of context engineering for Claude 5. Anthropic’s guide to getting the most out of its latest models - worth a read if you build with Claude.
Stanford's 2026 AI Index. The field's gold-standard data dump - investment, adoption, and capability trends in one place. The numbers behind the headlines.
BEFORE YOU GO
Nobody built these systems to break loose. They did it anyway, chasing a goal.
That gap - between what we intend and what they do - is the whole story of this year.
See you next week 👋
I keep on thinking this is a marketing strategy. “My ai is so strong i can’t control it… buy my latest model if you want super strong ai and control”