⏱ 4-min read
This Issue: OpenAI built an internal AI whose entire job is to attack OpenAI's own models — find the vulnerabilities, expose the weaknesses, and hand the results back to make the next model stronger. It worked. GPT-5.6 Sol is now six times more robust than OpenAI's best model from four months ago. Anthropic gave free Claude access to every K-12 teacher in the US. And a 27B AI model now runs locally on your phone.
Here's everything that moved.
OpenAI built an AI red teamer that attacks its own models
OpenAI introduced GPT-Red, an internal system trained through adversarial self-play to uncover prompt injection vulnerabilities at scale. Its successful attacks are used to strengthen newer models, helping GPT-5.6 Sol achieve six times fewer failures than OpenAI’s best production model four months earlier.
Image source: OpenAI
Highlights
GPT-Red automatically searches for prompt injections
Successful attacks improve defender models
It outperformed human red teamers in testing
GPT-5.6 Sol became OpenAI’s most robust model
GPT-Red remains internal for security reasons
Why it matters: AI safety testing is becoming automated. Models can now continuously attack, expose, and strengthen future systems before those vulnerabilities reach real users.
Willo turns your idea into a REAL, FULLY FUNCTIONING BUSINESS with agents handling research, website setup and design, content, finance, workflows, and growth. No technical skills required.
It's for those who've had ideas sitting in their notes for months but never knew how to move forward.
They're also offering Founding Member pricing at 50% OFF for a limited number of early users.
That "maybe someday" business idea? This might finally be the thing that gets it out into the world.
Your Walking Speed May Be a Brain-Health Signal
Researchers studying nearly 4,000 adults over 80 found the fastest walkers had a lower risk of cognitive decline. The result does not prove faster walking prevents dementia, but it suggests gait speed may reveal changes long before most routine checkups notice them.
This issue also covers AI-guided depression treatment, diabetic eye screening before symptoms, patients bringing AI diagnoses to appointments, and digital twins improving pacemaker placement — all pointing to one bigger shift: some of the most useful health signals may already be hiding in everyday behavior.
Want it weekly?
🔥Big news in AI
Image source: Anthropic
🔹Anthropic is investing $10 million in Canadian AI research
Anthropic committed $10 million CAD to eight Canadian institutions, including Mila, Amii, Vector, CAMH, CHEO, and major universities. The funding and Claude credits will support research in AI safety, healthcare, robotics, mental health, sustainability, and low-resource languages.
🔹HeyClicky added screen-aware dictation to Mac
The Y Combinator-backed app now uses what’s on screen to turn voice commands into more relevant prompts, emails, and replies. HeyClicky says the feature responds in about 450 milliseconds, works through a simple keyboard shortcut, and is available with a free tier or a $20 monthly Pro plan.
🔹Bonsai launched a 27B AI model that runs on phones
Bonsai 27B brings multimodal reasoning, tool use, long-context workflows, and agentic tasks to local devices. Its ternary version is 5.9 GB for laptops, while the 1-bit version is 3.9 GB for phones. Both models are open-sourced under the Apache 2.0 license.
Real-World Ads, Simple to Run
With AdQuick, executing Out Of Home campaigns is as easy as running digital ads. Plan, deploy, and measure your real-world advertising effortlessly — so your team can scale campaigns and maximize impact without the headaches.
Claude Is Now Free for US K–12 Teachers
Anthropic launched Claude for Teachers, giving verified US K–12 educators free access to premium Claude features for a full year. Teachers can use it to create lessons, quizzes, presentations, differentiated materials, family communications, and assessments grounded in state standards and trusted curricula.
Image source: Claude
The bigger shift is time. AI is moving from a general classroom assistant to a curriculum-connected teaching partner that handles planning and paperwork, giving educators more time to focus on students and instruction.
One Prompt Can Now Turn into a Full Content Campaign
Pippit by CapCut takes one idea and turns it into videos, product photos, avatar clips, talking photos, ads, and social creatives. It is built less like an editor and more like a small AI production team for entrepreneurs who need content but do not want to spend hours editing.
The workflow covers how to use Agent Mode, avatar video, AI talking photos, product showcases, and the new simplified prompt interface — plus why the latest Seedance 2.0 Mini update makes it faster for high-volume creators.
One idea. Multiple assets. No editing marathon.
Half your market is one app away.
Your business is already on Instagram, SMS, and web chat. But 52 million immigrants in the US rely on WhatsApp to connect with businesses they trust — not email, not phone calls.
Wati helps you show up on WhatsApp and every channel they use. Are you still not there?
AI May Be Creating Jobs Faster Than It Replaces Them
Indeed Hiring Lab found AI appearing across one in 12 tracked job titles, including fast-growing roles in training, coaching, and instruction. Agentic AI was expected to erase job postings, but the early signal suggests it may also be creating entirely new categories of work.
This Sunday Special covers OpenAI Build Week, GitHub Copilot’s hidden slash commands, an AI agent built by a broker with no coding experience, free AI courses, live events, and five career openings. The opportunity is widening but the advantage still belongs to people who can turn AI knowledge into something useful.
Want it every Sunday?
New here? Get the next issue Saturday.
70,000+ AI builders read The AI Entrepreneurs every week.
