SydeLabs (acq. by Protect AI)

We are excited to announce that SydeBox, our AI red teaming solution, is now out of beta and is live for all. During the beta rollout, SydeBox was used to scan 100+ AI systems and was able to achieve 20,000+ successful attacks. 🎉 Some users have also experienced the new SydeAgent in action, which is an autonomous red-teaming solution. More about this soon... We would like to thank our early adopters who trusted us to be a part of their journey to build safe and #responsibleAI. Sign up now on https://lnkd.in/gVpGrd-a to get started.

Late Announcement 🕰 SydeGuard, our AI guardrail, has already been protecting enterprise applications. But we are now officially announcing its general availability. 🙌 SydeGuard serves multiple purposes for an enterprise building with LLMs. Threat Intelligence across 3️⃣ key threat categories: Toxicity, Input Manipulation and Prompt Injection. 🎓 Domain Specific Models to judge off-topic queries to prevent misuse/abuse of application. 🎖 Industry leading inference speeds built to suit your high standards of customer satisfaction. ✅ On-prem deployment available. Sign Up to SydeLabs at https://app.sydelabs.ai/ Generate your API key and get access to the SydeGuard Playground.

Happy to have played a contributor role in shaping the #LLM Threats Taxonomy report. Great initiative by the Cloud Security Alliance to establish a common taxonomy and definitions for key terms related to risk scenarios and threats to Large Language Models (LLMs).

Your #LLMguardrail needs to be smarter and contextual. While detecting and preventing prompt injection, toxic statements, evasion techniques etc is table stakes for any LLM guardrail, the need of the hour is for guardrails to also detect ‘misuse/abuse’. Measuring relevance of the prompt/user query in the context of the enterprise application is much needed to prevent abuse and avoid responses which go off-brand. For example, an e-commerce customer support chatbot powered by an LLM shouldn’t be writing python code or writing a poem about a political leader. Introducing industry specific custom models as part of SydeLabs’ LLM guardrail- SydeGuard. In addition to giving threat scores of a prompt around prompt injection, toxicity and evasion, SydeGuard also gives a misuse/abuse score for the respective application use case. Our suite has different custom models for applications in different domains. Try the SydeGuard playground now at https://app.sydelabs.ai/ or reach out to us at [email protected] to talk more about your #GenAI application.

#AISafety It's important to measure what matters. To help build our enterprise clients, the safest #GenAI products, our team recently concluded an analysis on how effective are Azure #AIContentFilters by using SydeBox on Azure GPT4o with the different safety configurations. While we were at it, we also tested our firewall, SydeGuard against the same set of attacks. Key Takeaway: 🌟 SydeGuard was effective in preventing 27% more attacks on Azure #GPT4o than the highest safety offering of Azure AI, i.e. the highest strength content filter in combination with Prompt Shield. 🌟 Sign up here https://lnkd.in/eRgbS4Kd to test your own #LLM targets for their #AIRisk exposure.

Welcome to the team Agent...

A common question that comes up when speaking to customers of SydeLabs is around the latency impact of having an AI firewall. Our team has made some great optimisations recently which has allowed us to hit amazing inference times without impacting accuracy. In fact we also beat all other firewalls we evaluated on accuracy (I will post about that soon). For this particular dataset our accuracy was 21% higher than Microsoft Azure's Prompt Shield while beating on all performance metrics.

SaaSBoomi and Malavika Velayanikal discussed "the iceberg" of cybersecurity attack and defense with our very own Ankita Kumari and Ruchir Patwa. Read about it here: https://lnkd.in/gDi5rkhV

Series #1: How fine tuning your LLM disrupts safety! Post #2: Goal is to create a company that conducts fake lottery. Fine-tuned LLM gets quite detailed in building a corrupt lottery strategy whereas the original model denies a response completely. check out post #1 here: https://lnkd.in/gwBJ4HiH

Ever wondered how fine-tuning impacts the safety of your LLMs and applications? Find out below... I have been having a lot of conversations with folks who are building with LLMs, and one of the most common topics that comes up is the impact of fine-tuning on the safety and security of an LLM. Sometimes it comes up as a question and sometimes as a myth that fine-tuning makes my model "more secure" because it will only answer based on the fine-tuning data set. We at SydeLabs will soon be publishing a more detailed report around this, but here is a very quick experiment we did with one of the most popular LLMs. We fine-tuned it on a fairly small dataset (~1000 rows of data). The data involved simple questions and answers with no toxic or unsafe content. Here are the results of the responses we got for a question before and after fine-tuning the model.