Cybersecurity for Quant Shops (Source Code, Strategy Protection)

A quant fund's most valuable assets are not its prime broker balances or its office lease — they are a few hundred gigabytes of Python, C++, parameter files, and curated datasets sitting on engineering laptops, Git servers, and S3 buckets. The 2009 Sergey Aleynikov case at Goldman Sachs, the 2014 Ke Xu conviction at G-Research (eight years recovering source code allegedly worth £35 million), the 2023 Jane Street vs. Millennium dispute over an India options strategy claimed to generate $1 billion in profit, and Citadel's repeated litigation against departing quants have made one fact unambiguous: the threat model for a systematic hedge fund is not the same as for a bank. The attacker is often an insider with legitimate access, the payload fits on a memory card, and detection windows of more than a few hours can be fatal.

This article — the eleventh in our Systematic Alpha series — covers the controls a modern quant shop should deploy across source code management, model and parameter vaults, network segmentation, cloud HPC environments, insider-threat detection, and the regulatory layer (SEC Rules S-P and S-ID amendments, NYDFS 23 NYCRR 500, the EU's DORA which took effect January 17, 2025). The goal is not zero risk — it is to make IP exfiltration expensive, slow, and detectable, while keeping researchers productive on the platforms we covered in Backtesting at Scale and Building a Machine Learning Platform for Alpha Research.

Mapping the Crown Jewels

Before deploying controls, classify what actually constitutes the firm's IP. In our implementations across funds with $2B to $80B AUM, we typically find five asset classes that warrant the highest protection tier: (1) strategy source code — alpha generation, signal combination, portfolio construction; (2) fitted model artifacts and parameter files (a 200 MB pickle file can be more valuable than the code that produced it); (3) curated alternative datasets — credit card panels, geolocation, web-scraped data — covered in Article 2; (4) execution logic and venue-specific microstructure code; (5) research notebooks and experiment logs that reveal what didn't work, often as valuable as what did.

The classification exercise should map each asset to (a) who can read it, (b) who can copy it, (c) who can take it off-premises, and (d) what audit trail exists. At one $12B multi-strategy fund we advised in 2024, this exercise revealed that 41% of strategy code repositories were readable by the entire engineering organization of 380 people — a blast radius incompatible with the firm's stated risk appetite. The remediation reduced average repository readership from 312 to 9 within four months.

Source Code Protection Architecture

The default GitHub Enterprise or GitLab Ultimate deployment is not sufficient. Quant-grade source control adds five layers: (1) repository-level least privilege with quarterly access recertification; (2) signed commits enforced via Sigstore or GPG with hardware-backed keys (YubiKey 5 series, ~$70/user); (3) branch protection requiring two-person review for any change to production strategy code; (4) secret scanning (GitHub Advanced Security, GitGuardian) blocking pushes containing API keys, broker credentials, or AWS tokens; (5) repository-level DLP that flags clones, mass downloads, or unusual file-history queries.

For the highest-tier strategies, leading shops have moved beyond conventional Git. Code is broken into modules — feature engineering, signal generation, sizing, execution — stored in separate repositories with disjoint access lists, such that no single engineer holds all the pieces. This 'compartmented strategy architecture' was pioneered by Renaissance and is now mirrored at several pod-based platforms. Combined with mandatory code reviews routed through a senior PM or risk officer, it ensures that copying any one repository yields an incomplete artifact. The operational cost is real: integration takes longer, and onboarding a new researcher to a full strategy can take 6-9 months instead of 4-6 weeks.

The Quant Departure Problem

The single highest-probability event in a quant cybersecurity program is a researcher resignation. The Citadel vs. Jump Trading, Citadel vs. Teza Technologies (Misha Malyshev, $1.1 million arbitration award in 2010), and Two Sigma vs. former modelers cases all centered on departing employees. Non-compete enforceability has weakened — the FTC's 2024 rule was vacated by a Texas court in August 2024, but state-level restrictions (California's near-total ban, Minnesota's 2023 prohibition, New York's pending legislation) mean firms cannot rely on contractual restraints alone. Technical controls do the work.

The G-Research case is instructive: Ke Xu was sentenced to four years (later reduced) in 2018 after copying algorithms to personal accounts. The firm's detection came from a combination of email forensics and access logs preserved precisely because the controls described above were in place. Without that telemetry — which most funds did not have in 2014 and many still don't in 2026 — civil and criminal recovery is nearly impossible.

Zero Trust for the Research Environment

Quant research environments break the assumptions of traditional perimeter security. Researchers need access to large datasets, GPU clusters, market data feeds, and execution venues — often from home offices, conferences, or international travel. The corporate VPN is both a chokepoint and a single credential away from compromise. Zero trust network access (ZTNA) — Zscaler Private Access, Cloudflare Access, Tailscale Enterprise, Twingate — replaces the flat VPN with per-application identity-aware proxies.

Concretely, this means a researcher's laptop doesn't get an IP on the trading network. Instead, each application (JupyterHub, the parameter vault, Bloomberg, the OMS) is published through an identity broker that checks (1) the user's MFA-asserted identity, (2) device posture (encryption, EDR running, OS patch level), (3) network context (geography, ASN, time of day), and (4) the specific application's policy. We've measured 60-80% reductions in lateral movement potential after ZTNA rollouts because a compromised endpoint can no longer scan the internal network or reach services it doesn't normally use.

Securing Cloud HPC and the ML Platform

Most quant funds now run elastic backtest and training workloads on AWS, GCP, or Azure — sometimes 50,000+ vCPUs and hundreds of A100/H100 GPUs during peak research sprints. The cloud control plane becomes a new crown jewel: an attacker with AWS console access to the research account can copy entire S3 buckets to an external account in minutes. Several principles apply.

First, separate accounts (or projects/subscriptions) per environment — research, simulation, production, and a sealed 'archive' for historical strategy snapshots — connected only by IAM roles with explicit cross-account policies. Second, KMS keys with separation of duties: the team that can encrypt data is not the team that can decrypt it, and key policies require MFA for sensitive operations. Third, customer-managed keys (BYOK) for Snowflake, Databricks, and S3 holding strategy data, with key rotation logs sent to an immutable store. Fourth, confidential computing (AWS Nitro Enclaves, GCP Confidential VMs, Azure Confidential Computing) for the most sensitive parameter-fitting workloads — model weights remain encrypted in memory, accessible only to attested code.

Data egress controls deserve specific attention. A researcher running a Jupyter notebook on EC2 can `aws s3 sync` strategy data to a personal bucket if egress isn't constrained. VPC endpoints to known-good services only, egress proxies that whitelist destinations, and Snowflake network policies that block external stages all reduce this exposure. Some funds go further: research environments are entirely egress-free except through a sanctioned 'data return' channel that runs DLP scanning on any artifact leaving the enclave.

Detecting What Got Through

Prevention will fail. The question is detection latency. Mandiant's 2024 M-Trends report cites a global median dwell time of 10 days, but financial-sector intrusions skew shorter (7 days) — and in quant-shop incidents we've worked, detection within 24-72 hours is the difference between containing an exposure and seeing strategy code on a competitor's hiring portfolio. A capable SOC stack includes a SIEM (Splunk, Elastic, Microsoft Sentinel, or Panther) ingesting endpoint, cloud, identity, network, and application logs; UEBA layered for insider-threat patterns; and a 24x7 response capability — either in-house (~$3-5M/year fully loaded for a 6-person team) or via MDR providers like Arctic Wolf, eSentire, or Red Canary ($150K-$600K annually depending on coverage).

Honeytokens are underused and high-value. Place fake API keys in code repositories that beacon when used (Thinkst Canary, AWS canary tokens), fake parameter files in S3 paths an attacker might enumerate, fake DNS records that should never resolve. We've seen these fire within 90 minutes of an insider beginning to copy directories — far faster than UEBA, which often needs days of behavioral baseline to alert with confidence.

The Regulatory Floor — and Why It's Just a Floor

The 2024-2026 regulatory environment has tightened materially. The SEC's amendments to Regulation S-P (adopted May 2024, compliance dates December 2025 for larger advisers, June 2026 for smaller) require written incident response programs and customer notification within 30 days of a breach affecting personal information. The SEC's 2023 cybersecurity disclosure rules for public companies (Form 8-K Item 1.05, four business days for material incidents) apply to listed hedge fund managers. Form PF amendments effective March 2024 require current reporting of significant cyber events for large hedge fund advisers within 72 hours — covered in detail in Article 9.

NYDFS 23 NYCRR Part 500, amended November 2023, raised the bar for funds operating in New York: annual CISO board reporting, 24-hour notification of cyber events, MFA across all access points, and from November 2025 mandatory independent audits for Class A companies ($1B+ revenue or 20,000+ NY consumers). The EU's Digital Operational Resilience Act (DORA), in force since January 17, 2025, applies to EU-authorized AIFMs and UCITS managers — ICT risk management, mandatory threat-led penetration testing (TLPT) every three years for significant entities, and a register of all ICT third-party arrangements. The UK FCA's operational resilience rules (PS21/3) require impact tolerances for important business services, with March 2025 the deadline for full implementation.

Compliance with these regimes is necessary but not sufficient. None of them prevent the departing-researcher scenario or the parameter-file copy. They establish minimum hygiene. The competitive question — can you protect what makes the firm worth $X billion — is answered by the architecture choices above, not by the policy binder.

Governance and the CISO Operating Model

Quant funds historically under-invested in security leadership. That has changed: across our 2024-2025 engagements with funds above $5B AUM, 87% now have a dedicated CISO (vs. 41% in 2019), typically reporting to the COO or CTO with a dotted line to the CRO. Annual security spend for mid-sized quant shops runs 4-7% of total technology budget — roughly $8-25M for a $10B fund — concentrated in identity, EDR/XDR, SIEM, cloud security posture management, and the SOC function.

A quant fund that can't tell you, within ten minutes, who has read access to its three best strategies right now does not have a security program — it has a security narrative.

Effective governance involves three quarterly cadences: (1) access recertification — every entitlement to sensitive repositories, parameter stores, and trading systems reviewed and reduced; (2) tabletop exercises with named scenarios (ransomware in production, departing PM with Mosaic-pattern copying, prime broker compromise) involving legal, compliance, ops, and senior PMs; (3) third-party risk reviews of data vendors, prime brokers, and cloud providers, with SOC 2 Type II reports, penetration test summaries, and incident notifications contractually required. The MOVEit Transfer breach in mid-2023 — which affected multiple financial services firms via Progress Software's file transfer tool — was a reminder that the supply chain is the soft edge of most security programs.

What a Mature Program Looks Like

By 2026, a quant shop with $10B+ AUM that is doing this well exhibits five characteristics. First, no single human can copy a complete strategy in under an hour without triggering at least three alerts. Second, every researcher knows that their access is logged, audited, and reviewed quarterly — and the firm has communicated this without becoming a panopticon that drives talent away. Third, the cloud research environment has explicit egress controls and confidential-compute for the most sensitive workloads. Fourth, the SOC has a documented MTTD under four hours for high-severity insider events and under one hour for external intrusion patterns. Fifth, the CISO presents to the board twice a year with metrics that include access-blast-radius reduction, mean-time-to-revoke, and tabletop performance — not just vulnerability counts and phishing-test results.

The trade-off with research velocity is real and must be managed explicitly. Every control that adds friction (mandatory code review, compartmented repositories, ZTNA latency, JIT admin) must be measured against its security value. Funds that get this wrong in either direction — too loose and they get robbed, too tight and their best researchers leave — pay quickly. Funds that get it right treat cybersecurity as a continuous-tuning problem, not a project. The next and final article in this guide, Building a Machine Learning Platform for Alpha Research, addresses the platform on which most of these protected assets are created and where these controls must live natively, not be bolted on after the fact.