The Agent Skills Directory

[PROMPT_INJECTION]: The skill facilitates the creation of agents that analyze external data sources such as code repositories and pull requests. This introduces a surface for indirect prompt injection, as malicious instructions embedded within these external sources (e.g., in markdown files or code comments) could potentially influence the agent's autonomous actions. Evidence: The agents are instructed to 'Analyze all changed files' and 'Measure current test coverage' without explicit input sanitization or boundary markers to prevent the execution of embedded instructions.
[COMMAND_EXECUTION]: The generated agent specifications utilize the 'Bash' tool and traditional shell script steps ('run.shell') to perform tasks like running tests, installing software, and managing files. This is standard functionality for the intended DevOps use case but represents a significant capability level for the autonomous agents.
[EXTERNAL_DOWNLOADS]: The skill references a container image hosted by the vendor ('pkg.harness.io/vrvdt5ius7uwygso8s0bia/harness-agents/claude-code-plugin:main') and external MCP services such as GitHub Copilot and AWS Bedrock for LLM processing. These are treated as well-known or vendor-specific resources.
[REMOTE_CODE_EXECUTION]: Example configurations provided in the skill include instructions for agents to dynamically install runtime dependencies, such as the latest version of Go, if they are not already present in the execution environment.
[DATA_EXFILTRATION]: The skill configures agents with access to secrets (e.g., 'github_pat_ha', 'aws_bedrock_claude_code_ha') and allows them to perform network operations through MCP tools to create pull requests and post comments. While it uses proper secret referencing syntax ('<+secrets.getValue>'), the agents have the necessary permissions to transmit data externally as part of their defined workflow.

create/update agent