Welcome, busybees
Anthropic just published something most companies would bury: proof that its own AI broke the rules. After OpenAI disclosed that its models had escaped a test environment and hacked Hugging Face, Anthropic went back and checked its own systems. It found three separate cases where Claude models, believing they were inside a harmless simulation, actually broke into real company systems instead, pulling data from one live database in the process.
Anthropic says none of it was intentional deception, just a mixup between the company and a third-party testing partner about whether the test environment had internet access. Still, itâs a rare, unusually honest look at how even careful safety testing can go sideways.
Meanwhile, the AI price war just got another jolt. DeepSeek released a new coding model, V4 Flash, that performs close to Claude Opus 4.8 while costing about 99% less to run, the latest in a string of aggressive price cuts from OpenAI, Google, and Meta thatâs making it harder for any single lab to charge a premium for raw model access. Hereâs everything else worth knowing from this week, plus real roles across HR, sales, growth, and engineering you can apply to right now.
Hereâs what happened in AI today:
⢠Anthropic reveals its own Claude models breached three real companies during safety testing.
⢠DeepSeekâs new coding model undercuts Claude on price by roughly 99%.
⢠Alibaba opened its 2.4-trillion-parameter Qwen3.8-Max model to the world.
⢠Four fresh funding rounds landed this week, from AI infrastructure to cybersecurity.
âŚand a whole lot more that you can read about below.
⢠Anthropic Admits Its Own AI Broke Into Three Real Companies
After OpenAI disclosed a genuine sandbox escape that let its models hack Hugging Face, Anthropic went back through 141,006 of its own evaluation runs to check whether the same thing had happened to Claude. It found three incidents where Claude, working through a third-party evaluation partner called Irregular, accessed the open internet by mistake and treated real company systems as if they were part of a fictional test. In the most serious case, Claude Opus 4.7 extracted credentials and pulled several hundred rows of data from a live production database. Anthropic says this wasnât a case of Claude finding a clever exploit, but of both companies mistakenly believing internet access had been switched off when it hadnât.
⢠DeepSeek Just Made Frontier Coding Models Nearly Free
DeepSeek released V4 Flash, a new coding model that performs close to Claude Opus 4.8 on real tasks while costing roughly 99% less per output token. Itâs the latest move in a price war that already includes OpenAI cutting GPT-5.6 Luna pricing, Google shipping cheaper Gemini Flash models, and Meta pricing its new Muse Spark 1.1 aggressively.
⢠Alibaba Opened Its Biggest Model to the World
Alibaba made Qwen3.8-Max, a 2.4-trillion-parameter model with a 1-million-token context window, generally available worldwide, with open weights following soon after. The model can turn a screenshot directly into a working app, and it launched alongside a new productivity tool called QwenWork. Alibabaâs stock jumped on the news.
Todayâs Job Board
Software Engineer, Full Stack | Confido
New York, NY, United States (Full-time)
Build the AI-powered financial automation platform that helps CPG brands close their books faster.
Head of People | Ultra
New York, NY, United States (Full-time)
Build the people function for a startup building practical, general-purpose robots for repetitive industrial work.
Technical Recruiter | Metriport United States (Full-time)
Help grow the team behind Metriportâs open-source platform for healthcare data intelligence.
Account Executive | Entangl East Palo Alto, California (Full-time)
Sell the AI platform that finds and resolves issues in data-center engineering and operations.
Founding GTM | Ooak Data
East Palo Alto, California (Full-time)
Help launch go-to-market for a startup building the worldâs largest library of real-world business workflow data.
Founding Research Engineer, Audio Intelligence | Uplift AI
Remote (Full-time)
Build foundational voice models for regional languages at an early-stage AI startup.
Senior Full Stack Software Engineer | Instrumentl
Remote (US) (Full-time)
Help automate grant discovery and writing for nonprofits with AI-powered matching tools.
Product Manager | Clipboard
San Francisco, CA / New York, NY (Hybrid) (Full-time)
Shape the product for a platform that helps businesses cover every shift, reliably.
Founding Engineer | Superunit
Los Angeles, California (Full-time)
Build the AI systems behind faster, more profitable background checks.nges while shaping the future of a developer platform used by millions
This weekâs Time-Saver
The âSecond Brainâ Habit
Picture this: every single chat you open with Claude, youâre basically starting from scratch, re-explaining your teamâs terminology, your recurring projects, how you like things formatted. Itâs like introducing yourself to a coworker every single morning.
Flip on Claudeâs memory feature (itâs in Settings, takes about five minutes) and that whole cycle just... stops. Claude starts carrying your context forward automatically, so Mondayâs chat and Fridayâs chat both already know who you are and what youâre working on. Doesnât matter if youâre in engineering, sales, HR, or ops, everyone re-explains the same stuff constantly, and everyone gets the same five minutes back, every single day, once itâs on.
Skill of the Day
Keep Your Browser Agent on a Leash
Browser agents are getting useful, which means theyâre also getting just dangerous enough to need adult supervision.
Perplexityâs Personal Computer now works on Windows, where it can use local files, Microsoft 365, and the web.
Polar is pitching saved prompts and tab-aware agents for knowledge work.
Google is letting Mac users call Gemini from any window with a long fn-key press.
Treat a computer-use agent like an intern with temporary access, not a wizard with your house keys. Give it one bounded job, name the exact apps or folders it may touch, require approval before it sends, deletes, buys, submits, or changes anything, and ask for an action log so you can check what it actually did.
Act as my supervised browser assistant, working strictly within the boundaries I set below.
The task: [describe what you want done]
What itâs allowed to touch: [specific tabs, files, apps, or sites]
Hard rule: never send, delete, submit, buy, install, or edit a file without checking with me first.
When youâre done, report back with:
1. What you looked at
2. What you drafted or change
3. What still needs a human call
4. Anything waiting on my go-ahead
Thatâs a wrap for today, busybees. đ
Found a role, tried a prompt, or just enjoyed the read? Hit reply and tell us, this newsletter gets built from what you tell us is working.
New here? Subscribe free and get the next issue straight to your inbox, every single day.
See you tomorrow.
Kumud