OpenAI’s latest research shows how quickly AI can find security problems in software millions of people depend on.
But finding those flaws may soon be easier than fixing them.
Also in this issue:
Build your own AI assistant with n8n
Claude can now analyze your bank transactions
Grok Bot can make phone calls on your behalf
Sponsored
One inbox shouldn’t hold up your team
From AI and design to copywriting, SEO, and marketing, small teams use a lot of tools. And those subscriptions add up.
When teammates share the same premium subscription login, sign-in verification goes to one person’s inbox, leaving everyone else waiting for it to be forwarded.
That’s where HeyPass comes in.
HeyPass lets teammates request and receive email sign-in code/authentication in their own inbox, without the constant back-and-forth or sharing access to the original mailbox.
🎉 Try HeyPass FREE for 3 months. Use promo code 100NEW3 at checkout.
🔍 THE BIG READ
OpenAI’s AI exposed 948 Linux security risks
OpenAI introduced GPT-6 Cyber on Thursday, a new model built to find software weaknesses, test their impact, and help developers fix them.
And its early results are striking.
During a controlled security experiment, OpenAI’s automated research system:
Reproduced 1,724 vulnerabilities in the Linux operating system.
Developed 948 working exploits that could give someone with existing local access administrator-level control.
Spent roughly $1.2 million in API costs during the experiment.
These weren’t attacks on 948 live computers. The tests took place in controlled environments, and OpenAI developed proposed fixes for software maintainers to review.
Linux powers much of the internet’s infrastructure, making these findings especially significant.
The model also helped uncover 81 previously unknown high-severity flaws in Firefox. Mozilla has already released fixes for many of them.
But there’s another side to this progress.
AI tools that can discover security weaknesses this quickly could also be misused. OpenAI has classified GPT-6 Cyber as having critical cybersecurity capabilities and is limiting access to approved security teams.
The real concern is whether developers can fix these security flaws before someone uses them to cause harm.
See how OpenAI uncovered the Linux vulnerabilities →
🤖 AI IN ACTION
Build an AI assistant that answers questions from your own data
Learn how to use n8n to build a question-answering AI agent with its own knowledge base. The tutorial covers connecting information, setting up the agent, and testing its answers. Build your AI agent →
📡 ON THE RADAR
Claude can now look at your bank transactions: A new Personal Finance feature connects bank and card accounts through Plaid, helping users review spending and bills. It’s rolling out to US paid subscribers, with read-only access.
Perplexity joins the always-on AI assistant race: Its new Teammates agents can work on projects in the background, competing with OpenAI’s Dots, Anthropic’s Duet, Meta’s Muse, and Grok Bots.
Google’s AI video model now makes 15-second clips: Gemini Omni 1.2 Flash supports longer videos, follow-up edits, and more consistent characters. Google also reduced video-generation pricing by 30%.
Anthropic’s AI may be helping build better AI faster: Independent researchers found that Anthropic’s models are improving nearly twice as fast on measured benchmarks, with AI-assisted research likely driving much of the progress.
Cursor can now build and publish complete apps: Describe the app you want, and Cursor can handle the coding, database, user login, and publishing. You can then share the finished app with a link.
OpenAI’s GPT-6 Astra struggles with safety research: Internal tests found the model performed worse on safety-related coding tasks than similar tasks. The reason remains unclear.
🐦 TOP TWEET OF THE WEEK
Ring, ring. Your next caller might be an AI.
⚒️ WORTH A TRY AI TOOLS
Kino AI: Edit videos with AI agents such as Claude Code and Codex. Ask an agent to cut clips, add captions, or create motion graphics, then adjust the result on an editable timeline.
Momentic: Test websites and mobile apps using plain-English instructions. Its AI agents can run tests, identify problems, and help developers understand what went wrong.
We keep this spot for brands we believe in.
Interested in partnering? Reply to this email or reach us at hello@readaiblueprint.com
👋 BEFORE YOU GO
As AI takes on more responsibility, knowing when to ask a human for help could become one of its most useful skills.
Something to keep in mind as the next generation of assistants arrives.
Until next week.
948 working local-to-admin Linux exploits for about $1.2M in API spend turns CVE hunting into a budget line. The bottleneck moves to how fast maintainers can triage and ship fixes, because discovery no longer waits on a human reporter finding one bug at a time.