Chroma — agentic threat model

Not certain from the listing — Chroma is a vector database that integrates with external embedding models (like OpenAI or Hugging Face) rather than hosting its own foundation models, making L1 threats dependent on the chosen external provider.

Chroma is a vector database handling embeddings and metadata. Key threats include data/knowledge-base poisoning (injecting malicious embeddings to hijack downstream RAG outputs), embedding inversion (reconstructing sensitive raw text from stored embeddings), and unauthorized data exfiltration via query manipulation.

Not certain from the listing — Chroma provides the memory/retrieval layer for agent frameworks but does not orchestrate agent planning or tool execution itself; framework-level vulnerabilities depend on the orchestrator (e.g., LangChain, AutoGen) integrating Chroma.

Chroma can be deployed locally via SDK or hosted in a private-preview managed cloud. Threats include container/host compromise, unauthorized API access to the database port, and lack of network isolation between tenant namespaces in cloud deployments.

Not certain from the listing — Chroma focuses on storage and retrieval performance rather than built-in LLM evaluation or guardrails, meaning observability of retrieval quality or drift must be handled by external monitoring tools.

Not certain from the listing — As an open-source project with a private-preview cloud, built-in enterprise security controls (like RBAC, fine-grained access control, and compliance certifications) are not detailed in the basic listing and are typically self-managed in local deployments.

Not certain from the listing — Chroma acts as a shared memory resource but does not natively manage multi-agent coordination or marketplace interactions, though compromised shared databases can lead to cross-agent data contamination.