EQTY Lab

Agents need to drive fast without crashing. Today we're open-sourcing Cupcake: a policy enforcement layer for AI coding agents, with native integrations for Claude Code, Cursor, and more. github.com/eqtylab/cupcake

@HederaKimchi @Dell @hedera Thanks for the post!

📘 @Dell Releases @hedera Based @EQTYLab AI Verifiable Compute Whitepaper – Tokenization & LLM Validation Completed🧵 🔑 Executive Summary The Dell Trusted AI Verifiable Compute Whitepapergoes beyond traditional approaches that merely explain how AI computations occurred after the fact. Instead, it introduces a new standard for proving AI computations at the hardware level itself. Jointly designed by Dell, EQTY Lab, @nvidia, and @intel, Verifiable Compute performs AI training and inference within Trusted Execution Environments (TEE) across both CPU and GPU, while EQTY’s cryptographic AI Notary signs and records data, models, code, and execution environments in real time. Crucially, these cryptographic attestations can be registered on the Hedera Consensus Service (HCS), creating immutable, independently verifiable timestamped audit records. This architecture directly satisfies the technical requirements for trust, transparency, and accountabilitydemanded by regulatory frameworks such as the EU AI Act, sovereign cloud mandates, and compliance-heavy environments including financial institutions and the public sector. 📊 Benchmark & Performance Results (Chapter 5–6, p.21–25) 🔍 Core Validation Question: Does Trust, Security, and Verifiability Come at the Cost of Performance? Dell conducted comprehensive performance benchmarking to validate Verifiable Compute not as a proof-of-concept (PoC), but as a solution ready for real-world enterprise deployment. The benchmark environment was built on a standard enterprise AI infrastructure consisting of Dell PowerEdge R760 servers equipped with NVIDIA H100 GPUs. Importantly, the validation focus was not limited to raw compute speed, but rather measured actual performance with cryptographic proofs, TEE isolation, governance enforcement, and audit logging fully enabled. 🧪 Benchmark Design: Production-Grade Workloads 📌 Tested Workloads - Text Tokenization - LLM Inference (scaling from 3B to 70B parameter models) - End-to-end enterprise AI pipelines 📌 Test Conditions - CPU and GPU Trusted Execution Environments (TEE) enabled - EQTY Verifiable AI Notary applied - Real-time proof generation with full auditability ⚙️ Tokenization Results: 0% Performance Impact 🔹 Evaluated under large-scale text processing and data preprocessing workloads 🔹 Even with encryption, integrity validation, and cryptographic signing enabled, no reduction in throughput or increase in latency was observed ✅ Implications - Immediately applicable to foundational enterprise AI workloads such as data preprocessing, document analysis, and log processing - Definitively refutes the assumption that security and verification slow down baseline AI operations 🧠 LLM Inference Results: Average Overhead ≤ 4.5% 🔹 Across 3B, 8B, and 14B models, average performance overhead remained at or below 4.5% 🔹 Compared to traditional Confidential Compute environments, the incremental cost introduced by Verifiable Compute was minimal ✅ Implications - Real-time inference, customer-facing AI services, and analytics workloads can operate without perceptible performance degradation - Significantly lowers adoption barriers for AI in highly regulated industries requiring auditability and compliance 🚀 Large-Scale 70B Model Results: Overhead 0.1% 🔹 70B-parameter LLM benchmark - Performance overhead measured at 0.1% - Effectively negligible in real-world operations 🔹 Analysis - Performance overhead was primarily attributable to I/O and data movement, not computation - Once execution entered GPU memory and compute phases, performance was nearly identical to non-verifiable environments ✅ Implications - As model size increases, the relative cost of verifiability diminishes - Particularly well suited for institutional, financial, and nation-scale AI deployments operating large models 📊 Overall Performance Assessment - 📉 Average performance overhead: below 9% - 📉 Large-scale model overhead: 0.1% - 🔒 All measurements taken with security, governance, and audit controls fully enabled ✅ Conclusion Verifiable Compute empirically disproves the long-held assumption that enhanced security and trust necessarily degrade performance. The data clearly demonstrates that “Trusted AI does not have to be slow.” 🏁 Dell’s Message: Commercial Validation Complete Verifiable Compute is no longer an experimental or research-stage concept. It represents a commercially validated Trusted AI infrastructure that meets enterprise requirements for performance, security, and auditability. From tokenization workflows to large-scale LLM inference, Dell has demonstrated through measurable benchmarks that cryptographic proof generation and governance enforcement introduce only negligible overhead. This makes Verifiable Compute immediately deployable in regulated environments such as financial institutions, government agencies, public sector organizations, and global enterprises. Through this whitepaper, Dell positions Verifiable Compute not as a future vision, but as a practical, purchasable, and operable Trusted AI solution available today, backed by proven hardware, software, and enterprise services.

The Coalition for Secure AI (CoSAI), an OASIS Open Project, has released a blog post & white paper on MCP security: 12 threat categories, actionable controls, and defense-in-depth strategies for AI agent infrastructure.

If enterprises can’t perform AI agent governance, they can’t prove it. Proud to work with @EQTYLab to release Cupcake open source, and to have filed the original Hooks request and rationale with @AnthropicAI 🤝

Try Cupcake today: github.com/eqtylab/cupcake Docs: cupcake.eqtylab.io

Why OPA/Rego? Many agent security papers propose new policy languages. We built on Open Policy Agent instead—an industry standard your DevSecOps team already knows. Govern your AI stack with the same tools you use for everything else.

Agents need to drive fast without crashing. Today we're open-sourcing Cupcake: a policy enforcement layer for AI coding agents, with native integrations for Claude Code, Cursor, and more. github.com/eqtylab/cupcake

We're excited to join @FINOSFoundation in advancing the future of #VerifiableAI

Live from @RaiseSummit 2025: The #VerifiableCompute journey continues! For the first time, #VCOMP is available on @nvidia #Blackwell via @ScanComputers . Learn more about our collaboration with @Accenture Public Sector & @hedera eqtylab.io/blog/blackwell…

We’re excited to introduce a new AI solution with @Accenture and @EQTYLab that will help governments and public sector organizations manage AI-driven decisions with transparency, auditability, and trust via the @Hedera network 🤝

We're looking forward to collaborating with the @ConfidentialC2 community to advance verifiable AI systems 🚀

Ahead of this year's @RSAConference , @NSFOCUS_Intl profiled each of the top 10 Innovation Sandbox finalists. Thank you for the deep dive! #google_vignette" target="_blank" rel="nofollow noopener">securityboulevard.com/2025/04/rsac-2…

The next frontier of agentic AI is on silicon governance. Learn more about how EQTY Lab and @nvidia are leveraging confidential computing to enable the next leap in cybersecurity for AI agents. blogs.nvidia.com/blog/agentic-a…

Missed the new announcement from EQTY Lab at @RSAConference ? As a top 10 finalist in the #InnovationSandbox, founder Jonathan Dotan presented the Guardian, a new solution to govern AI agents. Check it out here: youtube.com/watch?v=OdwHlx…

We are proud to share that our portfolio company @EQTYLab has been named one of the Top 10 Finalists for the 20th annual RSAC™ 2025 Conference Innovation Sandbox contest for its work developing the first Verifiable Compute solution to align and audit AI agents at runtime. @EQTYLab will present its technology to a panel of renowned industry judges and a live in-person audience on Monday, April 28, at the RSAC 2025 Conference at the Moscone Center in #SanFrancisco. Full press release 👉eqtylab.io/blog/rsac-fina…

Next up: @EQTYLab, verified to trust #AI, on the #RSAC Innovation Sandbox stage

Listen in as we talk with our friends at @NEARProtocol during one of their AI Office Hours

Since @RiscZero’s zkVM runs verifiable computation, you’ll get an execution receipt alongside the Python program outputs. Hop over to our web verifier to check the receipt: risc0.verify.eqtylab.io

Following the release of our @RiscZero Verifier from yesterday, we are open sourcing another tool we built internally, this time for running Python code in the zkVM: github.com/eqtylab/pyzero

Our open source repo is at github.com/eqtylab/risc-z… It also includes a link to the hosted web verifier. Look for the “Try it out” section in the README for an example receipt that you can use.

We found this useful for increasing trust that ZK programs were run as expected so we’re open sourcing it. There’s a minimal verifier npm package, a customisable react component npm package, and a hosted version.

We’re fans of @RiscZero which allows code to be run verifiably. Instead of trusting a cloud provider, a zkVM provides cryptographic proof to prevent tampering during execution. We’re excited to release an open source web verifier that allows anyone to verify execution receipts.