Q12 of 60 — CCA Practice Exam

Multi-Agent Research System

Your multi-agent research system uses a code-analysis subagent to scan internal repositories and extract architectural patterns for research reports. The agent runs in a Docker container with the repository mounted via docker run -v /path/to/repo:/workspace:ro. Three weeks after deployment, a security audit flags that the agent's tool outputs occasionally include database connection strings and AWS access keys, which then appear verbatim in research report drafts. The :ro flag is confirmed to be set correctly. What is the most likely root cause and the appropriate fix?

A. The read-only flag prevents writes but not reads — credential files like .env, ~/.aws/credentials, and *-service-account.json inside the mounted directory remain fully readable by the agent. The fix is to exclude or sanitize those files before mounting, or copy only the required source files into the workspace, scoping what is mounted rather than just how it is mounted.

B. The subagent's tool descriptions are too permissive, allowing it to read any file rather than only source code files. The fix is to rewrite the tool description to instruct the agent to skip configuration files and focus exclusively on .py, .go, and .java source files, preventing it from opening credential files by instruction rather than by structural exclusion.

C. The research synthesis agent lacks a post-processing step to redact sensitive patterns before drafting the final report. The fix is to add a secrets-scanning pass over the synthesis agent's input that strips strings matching patterns like AWS key formats or connection string schemas before they reach the report draft.

D. The Docker image used by the agent has overly broad Linux capabilities that allow unrestricted file access across the container. The fix is to use --cap-drop=ALL when starting the container so the agent process cannot access files outside the workspace mount point, limiting its effective filesystem scope to only the intended directory.