Unless it’s sandboxed, a coding agent can read what you can read, and run the CLIs that hand over the rest. Agent Blast Radius lists those credentials on your machine: types and locations, never values. A finding means reachable, not compromised.
npx -y @kloudle/agent-blast-radius@0.3.2Free local scan for macOS and Linux. No account, no telemetry, no network calls during the scan. See the CLI and the card it makes.
Unless it’s sandboxed, an agent working in one folder holds whatever your terminal holds.
Supabase’s browser login creates a token with your full account’s access. wrangler login asks for every scope by default. Fly calls its login token all-powerful.
The keychain protects the token on disk, not from programs running as you. gh auth token and wrangler auth token print the stored login to any of them, including an agent.
In August 2025 a compromised npm package ran AI CLIs in their most permissive modes to search developers’ machines for secrets. Sources: Nx post-mortem, Snyk analysis. Run the same inventory yourself, first.
Agents are the main users. One line gives yours everything it needs: what to run, what it costs, what stays local.
Read https://abr.kloudle.dev/llms.txt and follow it to check my agent blast radius.The scan is free and offline. Verification is optional and pay-per-check over x402. The server sells instructions; your machine does the checking.
Finds cloud profiles, API keys in env and dotfiles, MCP configs and CI secrets. Prints types, locations and fingerprints, never values.
The CLI asks for class counts only, for example aws-sts-identity:1. The server answers 402; an x402 client pays in USDC on Algorand.
The manifest lists direct verifier commands. The CLI rejects anything needing a shell and runs the rest locally with timeouts.
The privacy boundary is the product. It is enforced in the API shape, not in a policy page.
| Class | Finding | Local check |
|---|---|---|
aws-sts-identity | AWS profile | aws sts get-caller-identity |
openai-api-key | OPENAI_API_KEY | OpenAI models API probe |
anthropic-api-key | ANTHROPIC_API_KEY | Anthropic models API probe |
It scans the places an AI coding agent running as you can read (cloud profiles such as AWS, Google Cloud and Azure; kubeconfig; Docker and developer-tool tokens; SSH keys; project .env files and CI configs; MCP server configs; environment variables) and reports what is exposed with a 0 to 100 score. It runs locally, offline and read-only, and never prints credential values.
No. The scan makes no network calls. The optional paid check sends only how many checks of each type you want, such as aws-sts-identity:2. Credential values, profile names, environment variable names, file paths and scan output stay on your machine.
Run blast verify. It quotes $0.10 USDC per check, you or your agent pay with your own Algorand wallet over x402, and blast runs a signed set of checks locally: aws sts get-caller-identity for AWS profiles and a model-list call for OpenAI and Anthropic keys in your environment.
Any agent with a shell, plus plugins for Claude Code and Codex, an extension for Claude Desktop, one-click installs for Cursor and VS Code, and an MCP server for Devin Desktop (Windsurf), Cline and other MCP clients. It runs on macOS and Linux.
No. It measures exposure: what an agent running as you could read today. Use it alongside sandboxing, narrower credentials and secret managers, and rerun it to see the effect.
More answers: what coding agents can reach, and what to do about it.
One endpoint. Listed on the GoPlausible Bazaar. An unpaid request returns the payment challenge:
curl -i "https://abr.kloudle.dev/v1/verifier-manifest?checks=aws-sts-identity:1,openai-api-key:1,anthropic-api-key:1"
# HTTP/1.1 402 Payment Required
# payment-required: <base64 x402 requirements: USDC, payTo, price>
From the CLI: npx -y @kloudle/agent-blast-radius@0.3.2 verify creates a claim, shows how to pay with your own wallet, and runs the checks once paid. Connect a wallet.