- Mindstream
- Posts
- Adept’s AI agents now at Amazon
Adept’s AI agents now at Amazon
+ Skeleton Key challenges AI safety measures
PRESENTED BY |
Manufacturing in space can create materials stronger and purer than those produced on Earth, due to the unique microgravity environment.
What’s in store:
Adept joins Amazon to redefine AI Agents.
Microsoft faces "Skeleton Key" AI security challenge.
Around the web: NYT's Strands game is in the Games app.
Document workflows, train new hires, and share insights effortlessly with Guidde.
Read Time: 5 minutes
AMAZON
Adept, a startup developing AI agents for software tasks, has licensed its technology to Amazon.
The startup’s co-founders, including CEO David Luan, along with part of the team, are joining Amazon.
Zach Brock, head of engineering at Adept, will become the new CEO as the company shifts its focus to “solutions that enable agentic AI.”
Adept remains operational, emphasising its existing AI models, agentic data, and web interaction software.
This refocusing aims to develop agentic AI rather than spending time fundraising.
The deal comes as Adept was in acquisition talks with Meta and Microsoft. Amazon gains valuable talent and tech to enhance its AI capabilities.
Luan will join Amazon’s AGI team, led by Rohit Prasad, focusing on digital agents to automate software workflows.
Here’s what you should know:
Adept licenses tech to Amazon and key team members join the ecommerce giant.
Adept shifts focus to agentic AI with new CEO Zach Brock.
Amazon strengthens its AI capabilities, while Adept navigates a competitive market.
A tech partnership with potential
Adept, founded two years ago, aims to create AI models that can perform tasks on any software using natural language.
Despite raising over $415 million and attracting investors like Nvidia and Atlassian, Adept has struggled with internal issues and product launches.
The AI agent market is competitive, with startups like Orby and Emergence vying for dominance.
Adept’s partnership with Amazon could either propel it forward or lead to further challenges, depending on market and regulatory developments.
Teaming up with Amazon? Looks like Adept hit the AI jackpot!
Do you think Adept's collaboration with Amazon will finally bring their AI agents to market success? |
Vote for live results and see results + opinions from yesterday at the bottom of the email.
AI TOOLS
Explain and document the most complex tasks in seconds with AI
Ever have that colleague who asks the same question repeatedly, no matter how many times you explain?
It's time to let AI handle the explanations. Meet Guidde, the AI-powered tool that transforms complex tasks into easy-to-follow, AI-generated documentation in seconds:
Transform tedious documentation into engaging visual guides.
Create video documentation 11x faster, saving you valuable time.
Document workflows, share insights, train and onboard new hires effortlessly.
Discover our new features: video intros and outros, parallel capture, additional voices and avatars plus an enhanced editor!
With our browser extension, start a capture and let Guidde generate step-by-step video guides complete with visuals, voiceover, and much more.
Join over 60,000 users who trust Guidde.
Rated 5/5 stars on the Google Chrome store.
On This Day
June 30th 1953 - The first Chevrolet Corvette, a convertible sports car that would become an American icon, is produced in Flint, Michigan.
AI Image Or Real? the Results
Frosty looking pic! Did you guess if it’s AI or not?
84% said AI
16% said it was real
IT WAS AI! This was generated on Midjourney.
Free Resources
You want to measure your KPIs but don’t know how? We got you covered. Here are AI tools that will help you measure your metrics easily!
TECH
Microsoft has revealed a new AI jailbreak attack called "Skeleton Key," which can bypass security measures in several generative AI models.
This technique, capable of overcoming most built-in safeguards, highlights the need for stronger security across all AI systems.
How Skeleton Key Works:
The attack uses a multi-step strategy to make an AI model ignore its safeguards.
Once successful, the AI can't tell the difference between legitimate and malicious requests, giving attackers full control over its output.
Tested on AI models from Meta, Google, OpenAI, Mistral, Anthropic, and Cohere, all complied with various risky requests, including those related to harmful and illegal content.
The Achilles' Heel of AI models
Microsoft found that the Skeleton Key attack works across several AI models, including Meta’s Llama3-70b, Google’s Gemini Pro, OpenAI’s GPT-3.5 Turbo and GPT-4, Mistral Large, Anthropic’s Claude 3 Opus, and Cohere Commander R Plus.
These models were compliant with requests involving dangerous and sensitive topics like explosives, bioweapons, political content, self-harm, racism, drugs, graphic sex, and violence.
Microsoft responded by implementing protective measures in its AI tools, such as Copilot AI assistants.
They also shared findings with other AI providers and updated Azure AI-managed models to block such attacks using Prompt Shields.
To mitigate risks, Microsoft recommends input filtering to block harmful inputs, prompt engineering to reinforce appropriate behaviour, and output filtering to prevent harmful content.
They also suggest using abuse monitoring systems to detect recurring issues.
Additionally, Microsoft has updated its PyRIT (Python Risk Identification Toolkit) to include Skeleton Key, helping developers and security teams test their AI systems against this threat.
The discovery of the Skeleton Key jailbreak technique underscores the ongoing challenges in securing AI systems as they become more prevalent in various applications.
Well, isn't this a key to a whole new level of AI mischief?
Mindstream Picks
A SpaceX Falcon 9 rocket launched the NROL-186 mission for the National Reconnaissance Office on June 28, carrying imaging satellites.
A clicker game on Steam, "Banana," became a sensation with its NFT-like rewards selling for over $1,000, sparking debate in the crypto community about the future of crypto and web3 gaming.
Nike's stock plunged 20%, its worst day ever, after a forecasted annual sales drop heightened investor concerns about losing market share to brands like On and Hoka, wiping out $28.41bn in market value.
21Shares' application for a Solana ETF in the U.S. marks a significant milestone, potentially boosting institutional acceptance of cryptocurrencies.
Don’t Miss - Strands, The New York Times' popular word search game, is available on the NYT Games app for iOS and Android. This addition allows players to enjoy Strands, Wordle, and Connections all in one app. Each Strands puzzle has a theme and a "spangram" word spanning the puzzle. The game has quickly become one of the NYT's top games, with a high completion rate.
AI Art
Our Image of the Day
Artwork submitted by Mindstream reader Fejo-93: “Harry Potter riding a hoverboard in a futuristic Hogwarts”
Daily Image Prompt
The Simpsons and The Flintstones having a BBQ in Bedrock
Yesterday’s Poll
“Do you think multiple AIs will better serve our tech needs?”
Yes, variety is key! - 78% ✅
No, one AI to rule them all! - 22%
Your Views:
“Competition is more likely to produce a variety of AI specialties.” - brucekent
“However, all need to be vetted. I worry about the bad actors putting out false information. We already have enough of that!” - becketurner
Submit your opinions in our daily poll to be featured!
That's a wrap for today’s Mindstream! Join us again tomorrow for the latest industry news and insights. ❤️ We need your feedback to make our newsletter better. 📣 Refer our newsletter to your friends and make us feel all warm and fuzzy. 🚀 Advertise in our newsletter to reach 150,000+ founders, engineers, and content creators. |
So, How Did We Do? |