Prem AI launches CyberScan for continuous AI security audits of Bitcoin infrastructure
London, UK, August 20th, 2026, Chainwire
CyberScan turns frontier AI models into repeatable, resumable security reviews that deliver actionable findings inside developers’ existing workflows.
Prem AI, the Sovereign Intelligence company with R&D and operations in Switzerland, today launched CyberScan, a continuous AI security audit harness built initially for Bitcoin infrastructure. Leading Bitcoin companies ArkLabs and Breez have had early access to CyberScan during the beta, running it against production code ahead of launch.
CyberScan gives engineering teams an ongoing layer of code review between formal security audits. It repeatedly analyses repositories as code evolves, preserves scan state if a run is interrupted, and returns prioritised findings with severity ratings and file-level references. It is designed to complement expert review, not replace it.
The recent wave of AI-driven attacks on Bitcoin projects made one thing clear: security reviews need to move at the same speed as the threats. We’ve been using CyberScan for the past few weeks, and it’s become a genuinely useful layer in how we review our code. Bitcoin infrastructure demands this level of rigor, and we’re glad Prem is building for it. — Roy Sheinfeld, co-founder and CEO, Breez
We got early access to CyberScan, and together with the Prem team we validated it against our own code, with our feedback helping shape the product along the way. Today, every release we ship goes through Cyberscan, and our code never leaves our control. — Marco Argentieri, CEO, ArkLabs
CyberScan’s launch follows the year’s largest hardware wallet exploit. On July 30, an attacker pulled more than $116 million in bitcoin from over 5,200 Coldcard hardware wallets, in four waves over a single week, according to TRM Labs.
The root cause was a firmware regression introduced in March 2021; human reviews on the open-source repository missed it, and automated AI attacks leveraged it: a build-time check meant to guarantee the wallet used its hardware random number generator instead silently let it fall back to a deterministic software PRNG. Every wallet seed became predictable, and every human review missed it for five years. The attacker rebuilt private keys offline, without ever touching a device.
Over the two weeks that followed, Prem AI served more than 50 companies across the ecosystem and processed more than 10 billion tokens, nearly all of it security research. The audits that came out of that work surfaced three critical vulnerabilities, each disclosed and remediated in coordination with the affected maintainers.
The response also exposed the limitations of model access alone. A model running in a chat session does not reliably preserve the state of a long review, resume after an interruption, or route its findings into the systems developers use to manage security work. CyberScan provides the workflow bundled in the harness.
Teams point CyberScan at a repository, where it maps the codebase, examines potential attack paths and produces structured findings with the relevant files and severity levels. Runs are resumable, while live progress and token usage make their status and cost visible. Teams can select the model used for each scan.
CyberScan is available through an API and Model Context Protocol (MCP), allowing teams to integrate it with their existing development tools. Findings are formatted for review and use in GitHub code-scanning alerts and issues, avoiding the need to extract results from a chat transcript.
Alongside CyberScan, Prem has launched a purpose-built version of DeepSeek V4 Flash 0731 that supports security workflows end-to-end and delivers state-of-the-art results on defensive security tasks. The weights are published openly on Hugging Face, and the model is selectable per run in CyberScan from day one.”
“Open-weight models have quietly become some of the best security reviewers in the world. Pair them with encrypted inference and they can do that work autonomously and continuously, on legacy code and new code alike, without a company ever giving up control of it. That combination is what CyberScan is built on. Bitcoin is the canary in the coal mine. It gets hit first, and the rest of financial services will follow.” — Simone Giacomelli, Founder and CEO, Prem AI
Prem is starting with Bitcoin because it is the hardest proving ground in security: high-stakes, open source, and under constant attack from opponents looking to turn a vulnerability into money. The same class of bugs that drained Coldcard lives across fintech and banking, and the company plans to bring CyberScan to the institutions running the most critical infrastructure on the planet.
Availability
CyberScan is available this week in beta at cyberscan.prem.io, with usage-based pricing and starter credits on signup. Model selection is per-run: Kimi K3 for high-stakes release audits, our purpose-built DeepSeek V4 Flash 0731 or Qwen 3.8 for high-throughput routine scanning.
About Prem AI
Prem AI is a Sovereign Intelligence company dedicated to delivering open access to frontier AI that is private, verifiable, and compounding. The company builds AI applications and infrastructure for organizations that manage sensitive information, integrating hardware attestation, confidential computing, and open-weight models so enterprises maintain absolute sovereignty over their AI. With R&D and operations in Switzerland, Prem serves institutions across the legal, healthcare, finance and government sectors.