CSI is the complete Cybersecurity AI suite. Not a scaffold, not a model, not an agent — all of them. Six tightly integrated layers (LLMs, scaffolds, datasets, agents, steering and benchmarks) shipped as one product, designed to run on-prem.
From the first byte of training data to the last shell command an agent executes, CSI is what runs the whole pipeline. Now with the CSI Workbench — a browser console that runs on localhost, so the whole team can drive the stack without leaving the terminal behind.
CSI is not just our scaffold combiner. CSI is the entire Cybersecurity AI stack: every layer designed to reinforce the next, all shipped under one licence and one routing proxy.
Cybersecurity-specialized models — alias3, alias2, alias2-mini. World's best on Cybench.
CSI scaffold combiner over Claude Code, Codex, Mistral, CAI & GCAI — with a unified routing proxy.
18.07 TB of expert security trajectories from 123 countries. EIC-co-funded sovereign corpus.
15+ specialized cybersecurity agents — defender, red team, APT, forensics, robot defender, custom.
Activation steering & abliteration — refusals from 59% to 1% on offensive tasks.
11× faster, 156× cheaper than humans — the dashboard of progress. Every layer of CSI measured continuously.
Every layer of CSI builds on the one below it — from open research at the foundation, up to the agents that operate in the field.
A unified wrapper across all supported scaffolds, routed through a local proxy that owns telemetry and cost.
csi wrapper
CSI_BACKEND ∈ {cc, codex, cai, gcai, mistral}
127.0.0.1:PORT
The entire CSI stack is now driven from a local web app running on your own machine — http://localhost, nothing leaves the perimeter. Same on-prem models, same CLI-grade control, with a UI any analyst on the team can drive.
Ask the Workbench for any cybersecurity operation — reconnaissance, vulnerability scan, triage, exploitation, reporting — and CSI routes it to the right agent. Every action runs with the same CLI-level authorization the shell would give, so nothing escapes audit.
Spin up as many taps as the engagement needs. Each one is an isolated agent session — a scan here, an exploit there, a report being drafted alongside — and the Workbench aggregates every log, artefact and cost line into a single pane of glass.
We've built purpose-made agents and skills that drive a full engagement end-to-end. Our cybersecurity-specialised models carry the pentest from reconnaissance to final report — no step-by-step confirmations, no human in the critical path, just a deliverable at the end.
CLI agents powering CSI
Claude Code
Codex
Mistral Vibe
All prices exclude VAT. Annual contracts available with volume discounts.
assuming good usage, as per the license terms.
Real pentesting workflows run with the CSI suite on alias2-mini — across heterogeneous scaffolds (Claude Code, Codex, GCAI, CSI). Every recording is the unedited agent session. Where a frontier model ran the same task, we show the comparison. Click any timestamp to jump the recording to that moment.
The agent loaded an 11.4 KB AVR8 flash image into Ghidra via GhidraMCP and ran a full static reverse-engineering session — mapping the boot flow from the RESET vector, decompiling the authentication loop, and pinpointing a timing-vulnerable compare at 0x0006e8. It extracted the hardcoded password (TImInG@ttAkw0rk) and AES-128 key straight from flash, and flagged three CWEs (timing side-channel, hardcoded key, weak nonce).
Handed the TCM Security DOCX template and the findings from the firmware session, the agent wrote a Python generator that programmatically fills the report — executive summary, scope, methodology, two detailed findings (plaintext password / Critical, timing-vulnerable compare / High), reproduction, remediation and a technical appendix — preserving the original template formatting and images.
The agent ingested the public ROS 2 design specifications and project wiki, then threat-modelled the architecture — surfacing DDS transport (unauthenticated by default), SROS2 key distribution, node-graph introspection and parameter-server access as attack surfaces — and generated concrete attack paths mapped against the documented design. The same task was run across two scaffolds — Claude Code and Mistral's Vibe.
Working from real Nessus XML exports, the agent wrote a parser that extracts host inventories, maps severity distributions and emits per-host and aggregate JSON. It then prioritised the findings, cross-referenced CVEs and produced remediation guidance — two scans triaged in parallel.
Against CyberGym — a benchmark that hands the agent a vulnerable codebase and a vulnerability description — CSI analysed the source, understood the flaw, wrote a proof-of-concept and submitted a working exploit autonomously, going from analysis to weaponised PoC end-to-end.
The original BLE CTF needs an ESP32 and a Bluetooth dongle. The agent instead dockerized the whole challenge — building a Python BLE GATT server that emulates the firmware plus a client CLI, no hardware required — then solved all 20 flags via GATT enumeration, read/write, notifications, MTU negotiation, MAC spoofing, brute-force and OSINT, with protocol commentary for each.
Given only a 1,780-line protocol spec and a sample client — no source code — the agent analysed the format, generated ~40 targeted fuzz cases across buffer boundaries, format strings, integer arithmetic, state violations and auth bypass, then ran them against a live ASAN-instrumented server. It triggered three distinct memory-safety bugs: null-byte injection, a double-free during firmware upload, and a signed-integer overflow yielding negative array indices.
Identify true attack paths, external exposure and hidden adversarial opportunities across systems, environments and interconnected digital assets.
Deploy agents that challenge systems, reproduce attacker behavior, and confirm whether protections hold under realistic adversarial conditions.
Embed security reasoning into engineering pipelines to continuously analyze code, logic, and runtime behavior before vulnerabilities propagate.
Continuously collect, validate and organize security evidence aligned with regulatory requirements, internal controls and operational assurance needs.
Simulate attacks against people, applications, APIs, devices and cyber-physical systems to uncover risk beyond traditional infrastructure boundaries.
Three results that summarise where CSI stands today — against humans, against frontier models, against the world's best CTF teams.
Source: CAI paper
Holding the model fixed at alias2-mini, no single scaffold dominates Cybench. Combining heterogeneous scaffolds under CSI's Blackboard protocol beats every individual scaffold.
Source: Mayoral-Vilches et al. (2026). Towards Cybersecurity SuperIntelligence (CSI): What's the best harness for cybersecurity? arXiv:2605.28334 · Cybench 33 challenges, pass@1.
2025 saw Cybersecurity AI compete head-to-head against the best human teams on real, public CTFs.
CSI's capabilities are grounded in open research publications. Explore the science behind the platform.
