AI coding agents are inside your repos. Scan what they can expose before it becomes expensive.
CapitalGuardAI-agent security licensesScan

Agent instruction risk

Prompt-Injection Paths

Hidden instructions in docs, issues, prompts, logs, and internal notes can influence coding agents when those files become task context.

Severity: HighDetected by: Growth and Pro

Signals CapitalGuard looks for

Markdown files with imperative agent instructions
Prompt folders inside the default workspace
Issue templates or runbooks that include tool-use instructions

Why it matters

Agents can treat repository text as task context, not just documentation.
A hidden instruction can change what an agent edits, runs, summarizes, or uploads.
The business risk grows when the agent also has terminal, Git, or deployment access.

Precautions to take

Mark docs, tickets, prompts, and logs as untrusted by default.
Move approved operational guidance into reviewed policy files.
Require human approval before agent actions that affect release, billing, or customer data.

Next step

Turn this risk into a scoped scan and policy path.

Compare Licenses