Lakera, an innovative Swiss startup, is on the forefront of developing safeguards for generative AI applications against pernicious prompts and other security risks, recently securing a $20 million investment in its Series A funding led by Atomico, a notable European venture capital entity.
The rise of generative AI, propelled by widely-used platforms such as ChatGPT, marks a significant leap forward in the AI field. However, its adoption in the corporate realm is hampered by ongoing concerns over security and the protection of sensitive information.
To put things in perspective, large language models (LLMs) fuel generative AI, enabling the technology to process and create text in a human-like manner. These applications require directives or ‘prompts’ to produce specific outcomes. Yet, these prompts can be manipulated to exploit the application in unintended ways, such as leaking confidential training data or breaching secure systems, through what are known as ‘prompt injections’. Addressing this escalating issue is Lakera’s core mission.
Responding to the Challenge
Originating from Zurich in 2021 and launching officially in October with an initial $10 million in funding, Lakera aims to shield organizations against vulnerabilities in LLMs, including data breaches and prompt injections. The startup’s technology is compatible with various LLMs including those from OpenAI, Google, Meta, and Anthropic.
At its essence, Lakera has introduced a novel “low-latency AI application firewall” designed to secure data flow to and from generative AI apps.
The flagship offering from the company, Lakera Guard, relies on a rich database aggregating diverse insights from open-source datasets found on platforms such as Hugging Face, proprietary machine learning research, and an ingenious game named Gandalf. This game challenges users to deceive the system to disclose a secret passphrase.
Through progressive levels of complexity, the Gandalf game has enabled Lakera to formulate a “prompt injection taxonomy”, categorizing various forms of these cyber attacks.
“Our approach is firmly grounded in AI. We’ve developed proprietary models that identify and neutralize prompt injections in real-time,” shared Lakera’s CEO and co-founder David Haber during an interview with TechCrunch. “These models are refining their understanding of malicious patterns through constant exposure to generative AI interactions, thus becoming progressively adept at countering emerging threats.”
Integrating with the Lakera Guard API allows enterprises to fortify their defense against hazardous prompts. Additionally, Lakera has pioneered specialized models for scanning both prompts and the outputs of applications for toxic content, equipped to detect hate speech, sexual content, violence, and crude language.
“These modules are invaluable for applications with a public interface, like chatbots, but are equally vital in other contexts,” stated Haber.
Like its security solutions, Lakera’s content moderation capabilities can be easily incorporated with minimal coding effort, supported by an intuitive policy control dashboard that allows for precise adjustment of content sensitivity settings.
With the additional capital of $20 million, Lakera is set to broaden its footprint globally, specifically targeting expansion in the U.S. market. The company boasts a roster of significant clients in North America, such as the AI innovator Respell and Canadian tech titan Cohere.
“A wide array of businesses, from large corporates to SaaS providers and AI technology developers, are eager to deploy secure AI applications,” Haber elaborated. “While the financial sector is at the forefront, acknowledging the criticality of security and regulatory compliance, interest spans across various industries. Most organizations recognize the imperative need to integrate generative AI into their operational workflows to remain competitive.”
In addition to Atomico, stakeholders in Lakera’s Series A funding round include Dropbox’s venture capital division, Citi Ventures, and Redalpine.
Compiled by Techarena.au.
Fanpage: TechArena.au
Watch more about AI – Artificial Intelligence
