cloudflare-worker-base
Fail
Audited by Gen Agent Trust Hub on Feb 15, 2026
Risk Level: HIGHPROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSCREDENTIALS_UNSAFE
Full Analysis
- Prompt Injection (HIGH): The skill is highly vulnerable to indirect prompt injection (Category 8).
- Ingestion points: The
cloudflare-debugagent ingests untrusted data viawrangler tail(logs) andcurlresponses. Thedeploycommand ingests code changes viagit diffto generate commit messages. - Boundary markers: No explicit delimiters or instructions to ignore embedded commands are used when processing this external data.
- Capability inventory: Multiple agents possess high-privilege capabilities:
wrangler deploy(production deployment incloudflare-deploy.md),git push(repository modification indeploy.md), andwrangler d1 execute(database manipulation ind1-migration.md). - Sanitization: No sanitization or validation of the ingested external content is performed before it is processed by the agent, allowing an attacker to potentially influence high-privilege actions.
- Credentials Unsafe (HIGH): The
cloudflare-debug.mdagent includes commands to explicitly list and identify secrets. - Evidence: Running
npx wrangler secret listandgrep -r "env." src/exposes secret names and their locations within the codebase. While the agent has a 'Do NOT' rule against exposing values, the exposure of keys and usage patterns to the model remains a high-risk data exposure (Category 2). - Command Execution (HIGH): Multiple components execute powerful CLI tools with significant side effects. The
cloudflare-deploy.mdagent anddeploy.mdcommand executewrangler deployto modify production environments. Thed1-migration.mdagent executes database commands, anddeploy.mdperforms git writes. - External Downloads (MEDIUM): The
init.mdandworker-scaffold.mdscripts performnpm installfor various packages. While these packages (Hono, Vite, Wrangler) are from reputable sources, the use of unverified, future-dated version numbers which do not currently exist is highly suspicious (Category 7). - Metadata Poisoning (MEDIUM): The README and configuration files reference software versions that are significantly ahead of the current stable releases (e.g., Wrangler 4.54.0 vs current v3.x, Vite 7.3.0 vs current v6.x). This misleading metadata could lead to unexpected behavior or reliance on hallucinated features.
Recommendations
- AI detected serious security threats
Audit Metadata