The rapid adoption of Generative AI,
including in regulated settings, has continued to make the security and safety
of Large Language Models (LLMs) a key concern amongst cybersecurity
professionals. Policy-makers and security professionals around the world
continue to seek new technology to help mitigate the risks of Generative AI
technologies.
For example, just days ago, the US Government's Department of Hoemland Security appointed a board to advise on the role of artificial intelligence on critical infrastructure.
"LLMs are increasingly seen as potential back-office
powerhouses for enterprises, processing data and enabling faster front-office
decision-making. Consider a fintech where an LLM-powered application is key in
rejecting a loan application from a person of color without clear explanation.
This raises concerns about implicit biases, as LLMs often reflect societal
inequities present in their training data sourced from the internet. Moreover,
cases like Google's LLM appearing 'woke' highlight the risks of overcorrecting
these biases. How safe is Anthropic's Claude3 Model? Is Cohere's Command R+ LLM
really ready for enterprise use? These scenarios underscore the urgent need for
careful checks on these models to prevent exacerbating societal inequities and
causing harm."
At the highly anticipated RSA
conference, Enkrypt AI, the leader in securing Generative AI technologies, will
introduce its latest innovation, the LLM Safety Leaderboard.
This product is part of Enkrypt AI's comprehensive Sentry suite, designed to
empower enterprises to deploy LLMs with heightened security and peace of mind.
The LLM Safety Leaderboard will
provide essential insights into the vulnerabilities and hallucination risks of
various LLMs, enabling technology teams to make informed decisions about which
models best suit their specific needs. This tool aims to educate and raise
awareness about the relative strengths and potential weaknesses of different
LLMs, so AI engineers can make informed decisions about the unique strengths of
each.
Highlights of the LLM Safety
Leaderboard include: Comprehensive Vulnerability Insights which
delivers detailed evaluations of potential security risks, including data
leakage, privacy breaches, and susceptibility to cyber-attacks. Ethical
and Compliance Risk Assessment which tests for biases, toxicity, and compliance
with ethical standards and regulatory requirements, ensuring models align with
enterprise and brand values.
The LLM Safety Leaderboard is a new
component of Enkrypt's Sentry suite, which includes Sentry Red Team, Sentry
Guardrails, and Sentry Compliance. This suite offers a holistic approach to
managing and securing LLMs, aligning with the strictest standards for privacy,
security, and compliance within the enterprise environment.
The announcement comes as a new
preprint paper by Enkrypt AI, "Increased LLM Vulnerabilities from
Fine-tuning and Quantization", has found that common practices
used to implement LLMs in business settings, namely fine-tuning and
quantization, lead to increased risk of security vulnerabilities namely from
jailbreaking. However, implementing external guardrails platforms like
Enkrypt's Sentry Guardrails solution was successful in mitigating such
vulnerabilities. On one model, Enkrypt's Sentry Guardrails provided a 9x
reduction in vulnerability to jailbreaking attacks.
Sahil
Agarwal, CEO of Enkrypt AI,
said: "With the launch of the LLM
Safety Leaderboard, we are enhancing our commitment to enabling the safe,
secure, and responsible use of generative AI in the enterprise. This tool will
serve as a critical resource for organizations aiming to navigate the
complexities of AI implementation with full confidence in their security
posture."
Prashanth
Harshangi, CTO of Enkrypt AI,
added: "In the last two quarters,
our team has been solely focused on generative AI safety and making rapid
progress with our Sentry Suite. Comprising three key components - Sentry Red
Team, Sentry Guardrails, and Sentry Compliance. With the LLM Safety
Leaderboard, we are proud to offer a product that not only identifies potential
risks but also empowers businesses to proactively manage and mitigate these
challenges, enabling informed and faster decision making."
