Risk 01
Unreviewed shell commands
Cline invokes bash to install, run, and test. One wrong command can drop a table, push to main, or fetch a remote payload. Without governance, there is no signed record of what ran and why.
Cline is an autonomous coding agent in VS Code. It reads your repo, runs shell commands, and writes code without asking. Raidu puts policy and proof around every action before it happens.
Cline (formerly Claude Dev) plans, executes, and revises code across files and the shell. Its autonomy is its value, and its governance problem.
Cline does not ask before it acts on most steps. That is the feature. It is also the audit question your CISO will raise the first time production changes.
Cline invokes bash to install, run, and test. One wrong command can drop a table, push to main, or fetch a remote payload. Without governance, there is no signed record of what ran and why.
To plan, Cline reads widely across the repo: config files, infrastructure as code, secrets. That context is sent to the model. Without redaction, it becomes a third-party record of your internals.
Cline edits files and stages changes. When those changes merge, the audit question is who reviewed the AI's plan, what policy applied, and whether the output passed safety scans.
Cline users configure raw OpenAI or Anthropic keys. Security loses per-repo control, policy routing, and revocation without key rotation.
Cline reads a base URL from settings. Point it at Raidu. Every plan, tool call, and response passes through the same runtime that governs Cursor, Claude Code, and your other agents.
Every file Cline opens, every shell output it captures, every instruction from the developer is scanned outbound. Secrets, PII, and flagged internals are replaced with deterministic tokens. The agent reasons with the tokens, not the raw data.
Shell commands match an allowlist. File writes match path globs. Destructive operations require approval. Network calls are scoped. The policy is shared with your other coding tools so the rules do not fork.
Model output is scanned for unsafe code, license risks, hallucinated packages, and exfiltration patterns before VS Code shows the diff. The developer sees clean suggestions. Blocked output is logged with a reason.
Each Cline turn becomes a signed record linked to the previous one. Prompts, tools, diffs, responses, policy version, user identity. RSA-4096 signed. SHA-256 chained. WORM persisted. Exportable for SOC 2 evidence.
Cline lets you pick the API provider and the base URL. Point both at Raidu and the agent is governed from that moment, across every project.
{
"cline.apiProvider": "openai",
"cline.openAiBaseUrl": "https://proxy.raidu.com/acme-corp/openai",
"cline.openAiApiKey": "raidu_xxx",
"cline.openAiModelId": "claude-sonnet-4.5",
"cline.customHeaders": {
"x-raidu-policy": "coding.eng.v7"
}
}
// Every Cline action now returns a signed record id in the response.