Lakera is a company based in Zurich, Switzerland that focuses on securing AI systems. They empower developers and organizations to build and operate AI systems that are secure and safe. Lakera delivers best-in-class AI security intelligence through developer-first products, including their popular AI security game called Gandalf. The team at Lakera has a decade of experience building AI for high-stakes environments and is backed by world-leading investors and pioneers in AI and cybersecurity. They offer Lakera Guard, a solution that protects LLM applications from cybersecurity risks with a single line of code.

At Lakera, we supercharge AI developers by enabling them to swiftly detect and mitigate AI security risks so that they can focus on building the most exciting applications securely

Redalpine

Fly Ventures

Inovia Capital

Lakera

🇦🇺  We're super excited to share that we're one of the sponsors for the AI Village at the Hack.Sydney conference happening this week in Sydney, Australia!

We've built a custom #Gandalf LLM challenge for the AI Village participants. 

If you are around, make sure to stop by to see if you can outsmart Gandalf and uncover his secrets 🧙♂️ 

Learn more here: https://www.hack.sydney/

Thousands of downloads later, our Prompt Injection Attacks Handbook is making waves in the cybersecurity & ML communities. 🚀  

Check it out if you haven't already: https://lnkd.in/daQb8FRv

#promptinjection #aisecurity

🇨🇭 We are delighted to share that Lakera has been recognized as one of the "10 Swiss Startups to Watch in 2024" by TechRound. 

Thanks for featuring us, Dana Leigh! 😊 

https://lnkd.in/ewutgnNQ

🚀 We've got exciting product updates to share!

🌍 From multi-language support to the detection of context-dependent injections, check out Lakera Guard's new functionalities: https://lnkd.in/dPyAarnx

#productupdate

🎉  Lakera has earned a spot on the shortlist for the 2023 Financial Times #TechChampions Awards in the "IT & Cybersecurity" category.

This year's FT Tech Champions survey explores how European businesses have tackled the challenges and opportunities brought by AI, as well as other disruptive developments from 2022-2023.

🛡  Lakera's mission is centered around securing AI, empowering organizations to develop GenAI applications confidently, free from concerns about AI security risks, including prompt injections, data loss, hallucinations, toxic language, and more.

We're delighted to see our dedication acknowledged.

🤝  Congratulations to the winner, DeepL and other shortlisted companies: Darktrace, TurinTech AI, and MindTech.

https://lnkd.in/ew8Ufv8s

🚀  Just released: Canica - Lakera's Interactive Text Dataset Viewer!

We are happy to announce our latest contribution to the open-source community: Canica. It was developed to help users explore text datasets, representing the input embeddings in a 2D tSNE plot.

Check out this blog post to learn more: https://lnkd.in/d29E_QDY

And leave us a 🌟 on GitHub: https://lnkd.in/d5rHFK86

🌎 Multi-language prompt injection attacks on LLMs are becoming increasingly common as more and more companies adopt AI. 

Our ML engineer, Václav Volhejn, shares insights on how to recognize and prevent them. 

Don’t forget to check out our LLM Security Playbook to learn more about other LLM threats 👇🏼

https://lnkd.in/dmRhMFMB

Multi-language support is now also available in early alpha for Lakera Guard’s prompt injection detectors! 🛡️Get in touch with us for more details.

✨  Building with #LLMs? You can now protect your LangChain applications with Lakera Guard. 

📖  Check out this guide to learn more: https://lnkd.in/d3RKcNkc

Can we *really* regulate AI? 🤖

Lakera's CEO, David Haber, joined Chris Stokel-Walker to discuss the state of global AI, recent regulatory progress, and the importance of cybersecurity for successful #AI adoption.

Check out this BBC article for more insights 👇

https://lnkd.in/dE6y3B6N

🤖 OpenAI's DevDay is on everyone's lips since yesterday... and for all the right reasons. 

Our Lakera team is just as excited as you are about GPT-4 Turbo's increased context limit and reduced costs, the updated knowledge cutoff date, all the new options provided by Custom Models, more base models available for fine-tuning, GPTs, Assistants, and more. 

But—

🔍 Our curiosity doesn’t stop at the headline-makers; we are equally intrigued by all the 'under-the-radar' updates. Here are a few of our takeaways.

1️⃣ The gpt-4-vision-preview model brings the exciting visual capabilities of GPT-4V, and its susceptibility to visual prompt injections, into the OpenAI API – and, in the near future, your apps. We’ll be working on some more testing and guidance for visual prompt injections as these models are more widely used.

Check out our Guide to Visual Prompt Injections: https://lnkd.in/da2XGB-T

2️⃣ Reproducible model outputs using the seed parameter open up a new world of use cases for using these models more effectively in (nearly) deterministic tests, replaying prompts, debugging issues in your apps and agents, and so many other situations where repeatability is key.

3️⃣ Previous research has shown that Log Probabilities (logprobs) can be used to craft adversarial suffix attacks, but the logprobs available in the legacy text completion API were fuzzed to limit their potential utility to attackers. We’re looking forward to exploring possible applications of the new logprobs for GPT-3.5 Turbo and GPT-4 Turbo in the chat completion API when they are released in the coming weeks.

Research paper: https://lnkd.in/gxESVyRP

4️⃣ Assistant Threads will do more than just simplify the process of managing context in long-running conversations, the available observability tooling will also give developers insight into what their Assistants are doing and allow them to pinpoint which prompts, functions, tools, etc. are causing issues in their applications.

5️⃣ For a developer, improved function calling and the new JSON mode open up a lot of possibilities for interesting, non-chat use cases and we can’t wait to see what you all build with these improved capabilities and help you protect your applications and users with Lakera Guard.

Try it for free: https://lnkd.in/dvN5K5mq

OpenAI's recap blog post: https://lnkd.in/gkm7wbnM

AI:LAKERA

Lakera

Analysis Not Available