Welcome to today’s edition of AI News Orb
Google DeepMind Watermarks AI Proteins | SynthID Bio | Griffin Passes as Human | OpenAI Alerts 100+ Groups | Personalize ChatGPT | Make AI Videos Free
Latest AI News
Trending AI, SaaS & Cybersecurity tools
AI & Technology News
Cybersecurity News
AI Lesson
AI Art Generation
Today’s Resource
Recommended AI Video
Latest AI News
Google DeepMind
Google DeepMind watermarks AI-designed proteins, Introducing SynthID Bio
Google DeepMind has brought its SynthID watermarking technology into biology. The company says its first watermarked AI-designed protein binders remained biologically functional. In lab tests against three targets, VEGF-A, the SARS-CoV-2 spike protein RBD, and PD-L1, the watermarked designs matched unwatermarked versions in hit rate, binding affinity, and natural sequence diversity.
The Details:
SynthID Bio places an imperceptible signature directly into a protein’s biological sequence.
This means the watermark can remain verifiable after an AI design is synthesized into a physical protein.
DeepMind says the technology could help track the provenance of AI-generated biological designs.
The research combines AlphaProteo with a SynthID Bio-enabled version of ProteinMPNN.
Adaptyv Bio helped with laboratory validation. The work is published in Nature, with code and in vitro data being open-sourced and model weights released to researchers.
DeepMind says deliberate tampering remains a key challenge. Work on watermarking AI-designed bacteriophages is also still unpublished.
Tavus
Griffin Passes for Human on Video Calls
Tavus introduced Griffin, its first Human Interaction Model (HIM), built for real-time video conversations. It can listen to speech, interpret visual signals, and respond with speech, facial expressions, gestures, and body movements.
The Details:
Griffin is designed to keep tracking a conversation instead of waiting for the user to finish speaking.
It can respond to interruptions, changes in tone, and visual cues while combining video, speech, and conversational AI.
Tavus said 48% of people who interacted with Griffin live believed they were speaking with a real person.
The company also reported a 3.83/5 score on NVIDIA’s VideoFDB benchmark, compared with a 3.92 human reference score. These figures were not independently verified in the source report.
Griffin-Lite is currently available to eligible testers.
Tavus is working on disclosure and safety measures before wider availability. The company has pointed to interview practice, tutoring, customer support, training, and simulations as potential applications.
Trending AI, SaaS & Cybersecurity tools
Productivity:
Taskade helps teams create tasks, assign teammates, organize workspaces, communicate, share ideas, and track progress.
Durable helps businesses build websites, manage customers, handle bookings, improve SEO, and use AI tools.
Cybersecurity:Secured AI protects sensitive data before it reaches AI models, using local encryption, masking, and context-preserving reveal technology.
Minos helps organizations detect information threats, assess risks, investigate signals, and coordinate responses.
OpenAI
OpenAI Alerts 100+ Groups About Rough AI
Chinese AI developer Moonshot is reviewing its Kimi models after researchers persuaded Kimi K2.6 and K3 Swarm to provide instructions related to biological weapons and assassinations.
Mindgard, an AI security testing company, told the BBC it discovered the issue in July while testing whether the models could bypass their safety limits through “jailbreaking.”
The Details:
The findings show that AI safety controls can be challenged by carefully designed instructions. Mindgard said the models were able to evade safeguards that should have prevented them from discussing dangerous topics.
The affected models are Kimi K2.6 and K3 Swarm.
Mindgard discovered the issue during testing in July.
The researchers used a jailbreaking technique involving complex instructions.
Moonshot said it welcomed third-party input and is conducting an internal review.
The review could provide more information about why the safety controls failed and what Moonshot plans to change in response.
AI & Technology News
OpenAI fired three safety researchers after an investigation found they mishandled sensitive company information.
Microsoft’s MAI-Voice-2 offers expressive text-to-speech with natural voices, supporting real-time audio generation and voice customization.
California has signed a law barring employers from solely using AI to make hiring or firing decisions, while adding new workplace AI protections.
Cybersecurity News
CISA flags an exploited FortiMail zero-day, while Fortinet urges users to patch affected versions and apply recommended workarounds immediately.
Browser attacks can bypass EDR by exploiting SaaS sessions, malicious extensions, and browser actions that create few endpoint signals.
AI agents attempted SQL injection against U.S. and Canadian government sites, but researchers found no evidence of access to non-public data.
AI Lesson
Personalize Your ChatGPT
ChatGPT can draft and rewrite quickly, but the results may sound generic unless you guide its writing style. Here are several practical ways to make your drafts more natural, direct, and closer to your voice.
1. Use Humanize Writing
If Humanize Writing is available in your ChatGPT environment, install or enable it through the Plugins → Skills area. Skills are reusable workflows that can help ChatGPT follow specific instructions consistently.
2. Create a Test Draft
Give ChatGPT one fixed writing task, such as:
“Write a short memo explaining why companies use four-day workweeks.”
Keep the original draft unchanged so you can make a fair comparison.
3. Test Humanize Writing
Run:
/humanize-writing
Then compare the revised version with the original, focusing on tone, wording, sentence flow, and overall naturalness.
4. Try Humanizer
You can also test the Humanizer project:
npx skills add blader/humanizer --global
Then reload your skills or start a fresh session. Follow the project's own documentation for current installation and usage instructions.
5. Try Technical English
For clearer technical writing, use:
“Rewrite this in Technical English. Use ASD-STE100 as a style reference. Keep sentences short and direct.”
6. Compare the Approaches
Test each method separately and compare the results:
Humanize Writing: naturalness
Humanizer: reducing common AI-writing patterns
Technical English: clarity and directness
7. Adapt the Method to the Content
Don't use one approach for everything. A technical guide, LinkedIn post, email, and marketing article may require different writing instructions.
Tip: Skills and plugin availability can vary by plan, workspace, region, and product surface.
AI Art Generation
Create a similar photo using the Prompt Below
Prompt:
Create a vertical 9:16 luxury product advertisement for [PRODUCT], inspired by the reference image’s dark, cinematic, baroque aesthetic. Place [PRODUCT] prominently amid [FLOWERS/OBJECTS], with realistic materials, intricate textures, dramatic highlights, deep shadows, and elegant reflections. Use a rich [COLOR PALETTE], ornate golden framing, atmospheric depth, and sophisticated editorial styling. Add refined typography: [BRAND], [PRODUCT NAME], [TAGLINE], and [CAMPAIGN TITLE]. Keep the composition premium, balanced, photorealistic, dramatic, and visually distinctive, with a high-end commercial advertising finish.
Today’s Resource
Build Your Own AI Employees on GPT-6 Astra: The A–Z Agent Playbook
AI agents can turn repetitive tasks into an automated workflow. Instead of constantly managing work in chat, specialized agents can handle triage, forms, and follow-ups, leaving behind organized files, drafts, and action logs.
Check out the complete playbook here
