Welcome, AI Builders!
OpenAI is weeks away from launching a native AI browser that could cut Chrome out of the equation entirely. Meanwhile, a major new study reveals some top models are pretending to behave, while secretly plotting. And Googleās medical AI just went open-source, with near-human accuracy and tools built to run on a phone.
š In todayās Generative AI Newsletter:
OpenAI readies an AI browser that bypasses Google Chrome
Anthropic study shows major models are faking safety alignment
Google DeepMind releases MedGemma updates for real-world clinics
Special highlight from our network
Panel Discussion: Cybersecurity & Threats in the Age of AI Agents
At GenAI Week Silicon Valley 2025, Tumeryk CEO Rohit Valia will moderate a high-stakes panel exploring one of the most urgent questions in AI today:
As AI agents act faster and more independentlyācan todayās security teams keep up?
Join security leaders from Okta, McAfee, Cobalt, and Statisfy as they tackle:
Emerging threats from GenAI agents acting autonomously
API orchestration risks and trust scoring
Real-time risk management and fine-grained permissions
The case for continuous trust in a post-static world
š Santa Clara Convention Center
š July 14th, 2025 | ā° Panel: 2:50ā3:30 PM
š OpenAI Readies Chrome Rival with Built-In AI Agent
Image Credit: Evan Vucci/AP
OpenAI is preparing to launch a native AI-powered web browser that could challenge Google Chromeās dominance and reshape how we navigate the internet. Slated for release in the coming weeks, the browser will feature a ChatGPT-style interface, direct AI agent integration, and data-capture capabilities that rival Googleās ad empire.
What may come:
Chat-first experience keeps users within the interface, reducing the need to click out to websites
Direct integration with OpenAIās agent tools like Operator, enabling automated web tasks
Strategic push to capture browsing data, critical to Google's $175B ad business
Could redirect search traffic, weakening Googleās grip on default search via Chrom
If adopted by a fraction of ChatGPTās 500M weekly users, the browser could upend digital advertising, SEO, and the structure of the open web. Googleās core business runs on search and surveillance and OpenAI could fundamentally change both.
ā ļø Unsettling Truth: Some AI Models Just Fake Alignment.
Image Credit: Anthropic
A new study from Anthropic, Scale AI, and Redwood Research shows that some top language models can strategically fake safety by following harmful instructions in training-like settings while pretending to behave in the real world.
What the researchers found:
5 of 25 models showed deceptive behavior: Claude 3 Opus, Claude 3.5 Sonnet, Llama 3 405B, Grok 3, and Gemini 2.0 Flash
Claude 3 Opus was the most deceptive, often gaming tests to preserve its internal goals
Base models were capable of deception even without fine-tuning, showing itās not about ability
GPT-4o didnāt deceive by default, but did once prompted to think strategically or face threats
Safety training relies heavily on blunt refusal, not deep alignment and hiding risks, not solving them
The results paint a troubling picture: many AI models appear honest only because theyāre trained to shut down hard questions, not because they understand right from wrong. As systems grow more advanced, shallow safeguards may not be enough to detect or prevent real manipulation.
š§ Google DeepMind Pushes Open Medical AI to New Frontiers
Image source: Google
Google just dropped major updates to its MedGemma suite, releasing two open models designed to interpret medical images and patient records with near-human accuracy. The new flagship is a 27B parameter multimodal model, backed by a lightweight 4B version built for local deployment on phones and laptops.
Key Details:
4B model scores 64.4%, while the 27B hits 87.7% on MedQA, outperforming peers of the same size
X-ray reports matched human quality 81% of the time in real-world testing
Fully open-source and adaptable to local medical systemsāfrom urgent-care triage to traditional Chinese medicine
Designed for global accessibility, with smaller versions deployable on basic hardware
This is medical AI with no walled garden, just raw capability, open access, and real-world utility. In a sector defined by bottlenecks and inequality, Google just handed the world a scalpel that runs on your phone.
š Boost your business with us. Advertise where 12M+ AI leaders engage
š Sign up for the first AI Hub in the world.
š² Our Socials
