CocoIndex is a data transformation framework that sits between raw data and the AI systems that depend on it. That position makes security a core requirement.
When developers trust CocoIndex to process sensitive documents, codebases, and enterprise data, the security of that path is our responsibility. We were selected to join the GitHub Secure Open Source Fund, a program that gave us structure and expert guidance to formalize our security practices.
What is GitHub’s Secure Open Source Fund?
When the Log4j “Log4Shell” vulnerability hit in December 2021, it showed that a bug in one small, underfunded open-source project can ripple through the software that depends on it. Most apps today rely on many dependencies, many maintained by unpaid volunteers, and a weakness in any one of them reaches everyone downstream.
Recognizing this, GitHub launched the Secure Open Source (SOS) Fund in late 2024 to strengthen the foundation of open-source security. The program pairs direct funding with a focused 3-week security sprint that combines expert guidance, better tooling, and a community of security-conscious maintainers.
In its first two cohorts, 125 maintainers from 71 projects participated, collectively fixing over 1,100 vulnerabilities, disclosing 50+ new CVEs, and preventing or resolving nearly 270 secret exposures.
- 71
- projects
- 125
- maintainers
- 1,100+
- vulnerabilities fixed
- 50+
- new CVEs disclosed
- ~270
- secret exposures prevented or resolved
Participants also exchanged knowledge, used AI-assisted tools like GitHub Copilot for security tasks, and developed incident response plans that many have since published for others to reuse.
What is CocoIndex?
CocoIndex is a declarative, Python-native data transformation engine for AI workloads.
As AI systems evolve from chatbots to autonomous agents, one hard problem keeps coming back: keeping context accurate and up to date as the underlying data changes.
Enterprises sit on large, changing datasets (documents, codebases, emails, APIs), yet most data infrastructure is still batch-oriented and blind to change. A change in data or logic usually means a rebuild from scratch, which breaks the feedback loop between the current state of the world and AI decision-making. CocoIndex keeps an AI system’s view of that data current and incremental: you declare the transformation, and the engine detects what changed and recomputes only the affected results. The model is close to React for UI, applied to data processing.
The cohort
We were in the program alongside Pandas, Apache Airflow, Fabric.js, PyPI, and others. We heard from teams dealing with similar challenges: how they handle dependency updates, structure security reviews, and think about trust boundaries.
Open source can be isolating. You’re deep in your own codebase, solving your own problems. Hearing how other maintainers answer the same security questions was one of the most useful parts of the program.
What we hardened during the sprint
During the program, the CocoIndex team ran a focused security sprint to harden both our codebase and the way we build and ship it. With guidance from GitHub Security Lab and community experts, we concentrated on four areas that matter for a framework sitting directly in AI data paths.
- 01Threat modelingPoisoned embeddings, prompt injection through source data, compromised dependencies
- 02Automated checksCodeQL data-flow analysis and secret scanning, Rust core and Python SDK
- 03Supply chainTransitive dependency audit, OpenSSF Scorecard checks, SBOM per release (in progress)
- 04DisclosureSECURITY.md, [email protected], triage and response commitments
Threat modeling and AI-specific risk assessment
We systematically threat-modeled the CocoIndex framework, examining how an attacker might exploit a data transformation pipeline that feeds directly into AI systems. We mapped out attack vectors including adversarial document content that could poison embeddings, prompt injection through source data that could manipulate LLM-based extraction, and compromised dependencies that could silently corrupt downstream indexes. Resources like the Adversarial AI Reading List and Embrace The Red were particularly useful in shaping this analysis. By identifying these failure points upfront, we developed mitigation strategies before they could be exploited in production pipelines.
Secure coding and automated checks
We wired GitHub’s security tooling directly into CocoIndex’s development loop. During the sprint, CodeQL ran on every pull request, scanning both our Rust core and Python SDK for vulnerability patterns (since October 2026 it scans main daily). For a system where data flows from sources through transformations to targets, CodeQL’s data flow analysis maps naturally onto our architecture, tracing untrusted input from ingestion all the way to exported artifacts. We also enabled secret scanning to keep credentials and API keys out of the repository, and we started using GitHub Copilot as a guardrail during development to surface potential issues before code is committed. CodeQL and secret scanning run automatically, so problems surface before a release.
Dependency and supply chain hardening
CocoIndex integrates with embedding models, vector databases, document parsers, and cloud storage backends, each one pulling in its own dependency chain. We audited our dependencies across both the Rust core and Python SDK, going beyond direct dependencies to examine the transitive tree underneath. We tightened our dependency review process with more systematic tracking of what we depend on and why. We also used the OpenSSF Scorecard checks to benchmark our project’s security posture across branch protection, dependency updates, CI/CD practices, and more. Following OpenSSF standards, we’re working toward a software bill of materials (SBOM) for each release, giving users visibility into exactly what ships with CocoIndex.
Policy, documentation and vulnerability response
We formalized our SECURITY.md with clear instructions for responsible disclosure and defined response commitments. Vulnerabilities should be reported to [email protected], not filed as public issues. We also established an internal process for triaging and patching reported vulnerabilities: who gets involved, how we communicate, and how quickly we ship fixes. Every new connector or transformation function now gets reviewed with two questions: what is the worst-case input, and what happens if a dependency in this path is compromised?
The automated checks catch classes of issues that a manual review can miss, and the disclosure process gives users a clear channel when something does slip through. Because CocoIndex sits in the data path of the AI systems built on it, these improvements carry downstream to the applications built on those pipelines.
Thanks to our contributors
This work rests on the people who build CocoIndex with us every day. Everyone who has submitted a PR, reported a bug, reviewed code, or helped another user in Discord: this project exists, and was recognized, because of you.
We especially want to thank those doing the work that doesn’t show up in commit logs: writing documentation, answering questions, and building examples that help new users get started.
What stays after the sprint
The sprint is over, but the checks it added stay in CI, and the review habits stay with us for every new connector and integration.
The GitHub Secure Open Source Fund gave us dedicated time, expert guidance, and a community of peers to learn from.
We’re grateful to GitHub and the entire cohort for this experience, and we plan to keep sharing what we’ve learned with the broader AI infrastructure community.
CocoIndex is open source on GitHub. Questions or feedback: join our Discord.
Frequently asked questions
What is the GitHub Secure Open Source Fund?
The Secure Open Source (SOS) Fund is a GitHub program launched in late 2024 to strengthen the security of open-source projects after the Log4j "Log4Shell" vulnerability showed how a single underfunded dependency can ripple through everything that depends on it. It pairs direct funding with a focused 3-week security sprint that combines expert guidance, better tooling, and a community of security-conscious maintainers.
Across its first two cohorts, 125 maintainers from 71 projects fixed over 1,100 vulnerabilities, disclosed 50+ new CVEs, and prevented or resolved nearly 270 secret exposures. See What is GitHub's Secure Open Source Fund?
What security improvements did CocoIndex make during the program?
CocoIndex ran a focused security sprint across four areas that matter for a framework sitting directly in AI data paths:
- Threat modeling & AI-specific risk assessment: mapping attack vectors like adversarial document content that could poison embeddings and prompt injection through source data.
- Secure coding & automated checks: CodeQL scanning (on every pull request during the sprint, daily on main since October 2026) plus secret scanning across the Rust core and Python SDK.
- Dependency & supply-chain hardening: auditing direct and transitive dependencies and benchmarking against the OpenSSF Scorecard checks.
- Policy, documentation & vulnerability response: a formal disclosure process.
How do I report a security vulnerability in CocoIndex?
Vulnerabilities should be reported to [email protected], not filed as public GitHub issues. CocoIndex formalized its SECURITY.md with clear instructions for responsible disclosure and defined response commitments, along with an internal process for triaging and patching reported issues.
How does CocoIndex use CodeQL for security scanning?
CocoIndex wired GitHub's security tooling directly into its development loop. CodeQL scans both the Rust core and the Python SDK for vulnerability patterns (on every pull request at the time of the sprint; since October 2026, a daily scan of main). Because data flows from sources through transformations to targets, CodeQL's data flow analysis maps naturally onto the architecture, tracing untrusted input from ingestion all the way to exported artifacts.
What is CocoIndex?
CocoIndex is a declarative, Python-native data transformation engine designed for AI workloads. You declare the transformation, and the engine detects what changed in the source data or logic and recomputes only the affected results, so the AI system's view of the data stays current without a rebuild from scratch.
What security risks are specific to AI data pipelines?
While threat-modeling the framework, CocoIndex mapped attack vectors specific to a pipeline that feeds directly into AI systems: adversarial document content that could poison embeddings, prompt injection through source data that could manipulate LLM-based extraction, and compromised dependencies that could silently corrupt downstream indexes.
Build fresh context for your agents.
CocoIndex keeps derived data in sync with its sources. Only what changed re-runs.


