Security by Design: How :Harvey: Engineered Trust from Day One
How we’ve built a proactive, layered security program to protect the most sensitive legal data—combining offensive security expertise, rigorous controls, and deep collaboration with our customers.
May 29, 2025
Multiple Authors
At Harvey, our goal is to build the most trustworthy AI platform for legal and professional services. From the moment we embarked on this journey, a core, non-negotiable principle has guided our every step: the security and confidentiality of our customers' data. We understand that the information entrusted to us is among the most sensitive and critical, and we’ve built our company from the ground up with this profound responsibility in mind.
A Culture of Security, Not an Afterthought
At Harvey, security has never been an afterthought. Our founders and first engineers laid a foundation of rigorous security standards, achieving both SOC 2 Type II attestation and ISO 27001 certification before we had a dedicated security team in place. Our early investments set the tone: security is not just a team – it’s a company-wide responsibility.
Our Head of Security joined as employee #23, a testament to our unwavering commitment from the outset. Today, our security team represents a significant portion of our engineering organization, hovering around 10-20% of staff dedicated to security.
Our Philosophy: Proactive, Offensive-Minded Defense
We view compliance as an outcome, not the driver, of our security program. Certifications like SOC 2 and ISO 27001 are important table stakes milestones in our industry, but they are the natural result of a robust security posture, not the ceiling of our ambitions.
Our security team brings a wealth of diverse experience, with many engineers having spent years in offensive security, as ethical hackers breaking into systems, or on the front lines of incident response, dealing with real-world breaches. This collective experience shapes our priorities, allowing us to focus on defending against the latest, most sophisticated threats.
We often go far beyond established practices. For instance, very early on, we implemented hardware tokens for multi-factor authentication (MFA) leveraging the FIDO2 standard. Unlike more common SMS, push, or time-code based MFA methods, FIDO2’s challenge-response protocol with digital signatures ensures that each authentication attempt is only valid for the website requesting it, and impossible to copy to other websites, effectively protecting against phishing attacks.
Partnership and Independent Validation: A Multi-Layered Assurance
Security is a collaborative effort at Harvey. We embed security engineers within each engineering team who build alongside their colleagues, providing dedicated security expertise and resources while maintaining independent reporting lines. This ensures we can innovate and ship features at unprecedented speed, without compromising on security.
Beyond our internal diligence, we actively seek independent, external validation of our security program by industry leaders:
- Rigorous Testing: We engage top-tier firms like NCC Group and Bishop Fox for regular application security assessments (penetration tests) and red team exercises. Penetration tests scrutinize our product for vulnerabilities, while red team exercises simulate real-world attacks on our corporate networks by tasking a team of ethical hackers to gain unauthorized access to customer data, often starting from an "assumed breach" scenario (e.g., a compromised engineer's or executive’s laptop). For us, these aren't just vulnerability hunts; they are crucial validations of our internal processes. Our goal is to prevent or detect any high-severity issues internally. If an external assessment uncovers such a finding, it triggers not only an urgent fix but also a comprehensive root-cause analysis and corrective action plan.
- Comprehensive Audits: We partner with Schellman, a firm renowned for its technical sophistication and attention to detail (the same auditors used by the likes of Microsoft, Zoom, and Salesforce), for our SOC 2 Type II and ISO 27001 audits, which are renewed annually.
- Customer Collaboration: We recognize that standards like SOC 2 and ISO 27001, while valuable, cannot capture every nuance, especially in the cutting-edge field of AI. We welcome and actively engage in deep, technical security reviews with our customers. Many of our largest clients, including many of the world's largest law firms, have conducted their own thorough assessments, and we pride ourselves on meeting their unique and stringent requirements.
Aggressive Third-Party Risk Management and Strong Contractual Commitments
We believe in providing transparency and peace of mind for our customers. Unlike providers who offer only point-in-time security information or require lengthy negotiations for substantive commitments, Harvey maintains strong, detailed contractual security commitments by default to all our customers. These include incident notification SLAs, vulnerability remediation SLAs, independent audits confirming our role as Data Processor, data residency controls, data deletion guarantees, defining the exact conditions under which our personnel can access your data, no training, and many more. Providing these commitments by default accelerates contracting and reduces redlining, allowing our customers to focus on the value Harvey delivers.
Our high security standards extend to our suppliers and vendors. We rigorously assess their security practices and make this a contractual commitment to our customers. This often means making tough choices and sometimes opting to build services in-house rather than engaging a vendor that doesn't meet our bar. As a leader in application-layer AI, we also leverage our position to secure enhanced commitments or tailored solutions from service providers, ensuring they meet our uniquely high bar for security.
Robust Governance: Ensuring Security Remains a Priority
Strong governance underpins our entire security program. Our Board of Directors is regularly briefed on emerging risks, the evolving threat landscape, audit outcomes, security events, and key metrics, such as the number of personnel with administrative access. We've also established an Information Security Risk Governance Council, enabling swift alignment and decision-making on security risks, reflecting our core company value of decisiveness.
Our Technical Approach: Layered Defense and Real-Time Response
Our security program is built on a layered approach:
- Strong security controls: We start with strong foundational controls that reduce complexity, aligning with another of our core values: simplicity. We thereby eliminate entire classes of risk and make the environment conform to specific patterns, making malicious activity easier to detect. To name a few of these controls: all data is encrypted at rest and in transit, developers do not have access to the production environment, and all development occurs in a strictly segregated environment devoid of Customer Data. We refer to this separation as "sacred." Any change to our systems requires two-party approval, with one approval from a designated owner for each code path or subsystem. We maintain a comprehensive audit trail of all changes and know precisely what code is running in our environment at any given time. A Kubernetes admission controller enforces that only authorized code can operate in our environment. Customer support personnel do not have default access to customer data; such access requires specific, individual authorization.
- Advanced Security Data Analytics: Building on this strong foundation, we've developed a highly scalable platform for real-time security data analytics. This allows us to ingest and analyze massive volumes of security signals in real-time, currently more than 2 billion per day. This platform enables us to create high-fidelity alerts tailored to our unique environment and business context, facilitating real-time threat detection, hunting, and investigations.
- Rapid Incident Response: We maintain a 24/7 incident response capability with flexible processes, enabling us to respond quickly and accurately to a wide range of malicious activities, and we contractually commit to notify customers within at least 48 hours for Security Incidents.
Addressing AI-Specific Risks By Design
We designed Harvey from the start to defend against risks that are unique to AI applications, such as prompt injection, sensitive information disclosure, training data poisoning, and excessive agency. If models are trained on confidential data and then shared with other customers or users, there is a risk that this data could be inadvertently disclosed to unauthorized parties. At Harvey, we do not train or fine-tune models on your sensitive data and the same applies for our subprocessors and model providers. Instead, customer data is used at inference time only, categorically eliminating the risk of sensitive data becoming part of the model and thereby categorically protecting against sensitive information disclosure risks.
Attack vectors such as prompt injection and training data poisoning could lead to the system producing incorrect outputs by processing maliciously-crafted documents. While we and our subprocessors have implemented defenses against these types of attacks, no large language model today is completely immune. However, Harvey is designed to assist lawyers, not to replace them. Our product is built to make verification easy, reducing the risks of errors or manipulation and aligning with the high standards expected of legal work.
Our Track Record: Demonstrable Security Excellence
Our commitment to security is reflected in our achievements:
- Sustained Certifications: We hold SOC 2 Type II attestation and ISO 27001 certification from Schellman, renewed annually.
- Proven Resilience: We successfully withstood an intensive two-week assumed-breach red team exercise conducted by Bishop Fox.
- Customer Trust: We have not failed a single security assessment from our customers, which now include a majority of the largest global law firms and a growing roster of major global enterprises.
- Data Privacy Leadership: Harvey was the first AI/LLM startup to certify under the EU-US Data Privacy Framework.
Job’s Not Finished
The landscape of security risks is constantly evolving with new technologies, attack patterns, and the rapid growth of our own company and products. While we are confident in our current security posture and the robust culture and structures we’ve built around it, we remain vigilant. Protecting customer data isn’t a one-time checkbox; it’s an ongoing commitment. That’s why we’re relentlessly focused on anticipating risks, adapting our defenses, and investing in long-term security.
We understand the trust our customers place in Harvey, and we are dedicated to earning and maintaining that trust every single day. Security is not just a feature at Harvey; it is our promise.