Dropbox needed to secure a rapidly expanding set of GenAI features—such as intelligent search, document summarization, chat interfaces, and automated content analysis—without slowing down collaboration or changing its self-hosted architecture. To do this, Dropbox selected Check Point AI Agent Security as its enterprise AI security platform after evaluating several vendors and internal options. The decision was driven by three main capabilities:

• Centralized, real-time protection: Instead of adding separate security controls to each AI use case, Dropbox integrated Check Point AI Agent Security directly with its centralized LLM library. This created a single security layer for all GenAI traffic across products.
• Ultra-low latency: Through close engineering collaboration, Dropbox and Check Point achieved a 7x latency improvement for extended-context prompts over 8,000 characters. After optimization, security analysis still maintained sub-500 ms response times, so users did not notice added delay.
• On-premises deployment: The solution was deployed as a containerized Docker service in Dropbox’s Kubernetes environment and accessed via internal RPC. This kept all customer data inside Dropbox’s own infrastructure, supporting privacy and regulatory requirements.

The result: Dropbox established a scalable AI security foundation that protects GenAI applications by default, while preserving the real-time experience that underpins its competitive position.

As Dropbox expanded its GenAI capabilities, the security team focused on emerging threats that traditional cybersecurity tools were not designed to handle. The main concerns included:

• Prompt injection attacks: Malicious users craft inputs intended to manipulate LLM behavior, bypass safety guardrails, or exfiltrate sensitive data. These attacks pose risks to data confidentiality and system integrity.
• Jailbreak attempts: Inputs designed to override model alignment and safety protocols, potentially enabling unauthorized access to sensitive information or generating harmful content.

Check Point AI Agent Security addresses these risks through a layered, defense-in-depth approach:

• Dropbox engineers built multi-stage “security chains” using LangChain, routing user prompts through Check Point’s prompt-injection and jailbreak-detection APIs before they reach the language models.
• LLM responses are then passed through content moderation analysis before being returned to users.
• The platform uses ML models trained on adversarial AI attack patterns and provides granular confidence scores, allowing Dropbox to tune sensitivity to its risk tolerance.

In testing and production, the platform achieved:

• Over 98% detection rates for prompt injection and jailbreak attempts.
• A false positive rate below 0.5%, which is important to avoid blocking legitimate user queries.

This combination of high detection accuracy and low false positives helped Dropbox strengthen system integrity and security posture across all GenAI touchpoints without adding unnecessary friction for users.

Before centralizing AI security, individual Dropbox product teams had to handle their own security reviews and custom protections for new AI features. This created bottlenecks and slowed feature launches. By integrating Check Point AI Agent Security as a shared, centralized service, Dropbox was able to reimagine how AI security fits into its development process:

• Security by default for all GenAI traffic: Because the platform sits in front of Dropbox’s centralized LLM library, any new GenAI feature—whether RAG systems, document Q&A, or text summarization—automatically benefits from the same protection, monitoring, and policies.
• Faster feature delivery: With guardrails already in place, product teams can focus on user experience and functionality instead of rebuilding security controls for each use case. This removed a previous bottleneck where security reviews delayed launches.
• Centralized visibility and nuanced responses: Security teams now see attack patterns across all products from a single vantage point. Using confidence scores from Check Point’s APIs, Dropbox can:
  • Immediately block high-confidence threats,
  • Flag medium-confidence inputs for human review and logging, and
  • Allow low-risk queries to proceed without friction.

According to Dropbox’s security engineering leadership, this centralized model has given the company the security foundation it needs to scale GenAI confidently across all product lines—without turning security into a bottleneck or compromising the user experience.