The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it manages pipelines that ingest data from external sources like Kafka, Kinesis, and S3. Maliciously crafted data from these sources could attempt to influence the agent's logic or the downstream processing pipeline.
Ingestion points: Data enters the system through $source stages for Kafka topics, Kinesis streams, and MongoDB change streams as described in SKILL.md and references/pipeline-patterns.md.
Boundary markers: The skill uses structured JSON pipeline definitions, providing structural separation between data and instructions.
Capability inventory: The atlas-streams-build tool can create processors with powerful capabilities, including database writes ($merge), outbound network requests ($https), and AWS Lambda invocation ($externalFunction).
Sanitization: The instructions recommend mitigation strategies, including the use of $validate stages for schema enforcement and Dead Letter Queues (DLQ) for isolating problematic documents.
[REMOTE_CODE_EXECUTION]: The skill supports the creation of processors that execute custom JavaScript User Defined Functions (UDFs) via the $function stage. While this is a native feature of the MongoDB Atlas Stream Processing service, it facilitates the execution of runtime-defined logic.
Evidence: references/pipeline-patterns.md details the $function stage for custom JavaScript logic, noting it requires an SP30+ tier.
[EXTERNAL_DOWNLOADS]: The skill references external resources from a well-known service provider to assist the agent in pipeline construction.
Evidence: SKILL.md directs the agent to fetch pipeline patterns and quickstart configurations from the official MongoDB Atlas Stream Processing examples repository at github.com/mongodb/ASP_example.
[DATA_EXFILTRATION]: The skill enables the configuration of data sinks that can send data to external endpoints, such as $https webhooks and $externalFunction (AWS Lambda). If an attacker successfully performs an indirect prompt injection, these sinks could be leveraged to exfiltrate sensitive data retrieved from databases during the enrichment phase.
Evidence: Sink patterns for $https and $externalFunction (async) are documented in references/pipeline-patterns.md.

mongodb-atlas-stream-processing