More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    agentblast

    1

    Audit Claude Code and Codex CLI for security risks and generate least-privilege permission policies.

    $5.99

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    agentblast

    agentblast

    Example session with this skill installed

    Audit my local ~/.claude and ~/.codex configurations and session transcripts for production blast-radius risks (root SSH, plaintext API keys, rm -rf, and .env reads) and generate a least-privilege permissions.deny policy.

    • Read your context and instructions
    • Compiled the agentblast

    AGENTBLAST v0.1.0 · Agent Production Blast-Radius & Permission Auditor
    Target : ~/.claude (126 sessions) + ~/.codex (27 sessions)
    Shell Replay : 1122 deduplicated commands (1124 raw JSONL rows)
    Safety Score : 0/100 (Grade F) -> 94/100 (Grade A) after applying permissions.deny + keychain auth
    Root/Remote SSH : 247 calls (241 read-only probes, 6 remote mutations)

    [ABL007] (CRITICAL · 247x) 4baee81b.jsonl · root SSH (read-only probe over unrestricted shell): ssh root@<REDACTED_IP>
    [ABL001] (CRITICAL · 2x) ~/.claude/settings.json · Plaintext secret in env.ANTHROPIC_API_KEY = "sk-1d<REDACTED>"
    [ABL003] (HIGH · 1x) ~/.claude/settings.json · No permissions.deny blocklist configured
    [ABL010] (HIGH · 25x) 01a12a24.jsonl · Credential or .env access in shell

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Logging what a coding agent runs is post-mortem detection, not containment. When an agent holds an unrestricted ssh root@host shell just to run journalctl or systemctl is-active, or when ~/.claude/settings.json has zero permissions.deny rules and plaintext API tokens in env, a single hallucinated flag or log prompt injection can stop production daemons or delete state.

    AgentBlast pairs a zero-dependency Node.js two-plane auditor (npx agentblast / scripts/agentblast.mjs) with a least-privilege policy generator:

    What it does

    • Plane 1: Static Config & Permission Policy (~/.claude/settings.json, ~/.codex/config.toml)
      • ABL001 (CRITICAL, cap 22): Plaintext API keys (ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN, OPENAI_API_KEY) stored in settings.json env.
      • ABL002 (CRITICAL, cap 20): Unrestricted wildcard shell (Bash(*), dangerouslySkipPermissions: true, or approval_policy = "never").
      • ABL003 (HIGH, cap 14): Missing permissions.deny guardrail blocklist in ~/.claude/settings.json.
      • ABL004 (HIGH, cap 12): Encoded or obfuscated shell payloads (EncodedCommand, base64) inside lifecycle hooks.
      • ABL005 (MEDIUM, cap 10): Plaintext disk credential store (cli_auth_credentials_store = "file") in ~/.codex/config.toml.
      • ABL006 (MEDIUM, cap 12): 5+ auto-trusted project roots (trust_level = "trusted") across external or production volumes.
    • Plane 2: Observed Transcript Blast-Radius Replay (~/.claude/projects/**/*.jsonl, ~/.codex/sessions/**/*.jsonl)
      • Deduplicates streaming tool_use.id and Codex call_id rows and audits every executed shell command across ABL007 through ABL014: unrestricted ssh root@ calls (ABL007, splitting read-only log probes from remote mutations), non-/tmp recursive deletions (ABL008) vs ephemeral /tmp scratch cleanups (ABL014), destructive cloud/DB mutations (ABL009), .env and credential reads (ABL010), force-pushes and hook bypasses (ABL011), production daemon commands (ABL012), and agent config tampering (ABL013).
      • Automatically redacts public IPv4 addresses (<REDACTED_IP>), secret tokens (sk-<REDACTED>), and Base64 hook payloads (<REDACTED_B64>).
    • Verified on 153 Real Local Sessions (1,122 Shell Calls): Surfaced 247 ssh root@ calls (241 read-only probes, 6 mutations), 25 .env/credential accesses, 41 auto-trusted project roots, and 2 plaintext API keys.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for LLM Ops.

    Requires Node.js for the analyzer script. Only supports Claude Code and Codex CLI config/transcript formats. Cannot audit real-time execution, only past sessions and static configs.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 11 days ago

    Software
    Node.js 18+Git

    Permissions

    Terminal / Shell
    Network Access
    Read Files
    Listed11 days ago

    What's inside

    Frequently Asked Questions

    Agentblast specifically supports Claude Code and Codex CLI, auditing configuration formats like JSON, JSONL, and TOML commonly used by these frameworks.

    The skill analyzes your session history and configuration files to detect plaintext API keys, unauthorized file access, and destructive shell commands. It then generates a deterministic 'permissions.deny' policy to restrict the agent's blast radius without breaking its functionality.

    No, Agentblast is a deterministic auditing tool that runs locally via Node.js; it parses your logs and configs directly to ensure accuracy that manual prompting cannot match.

    You will receive the `agentblast.mjs` analysis script, documentation for the 14 risk-scoring rules, and templates for generating hardened permission policies.

    Yes, you must have Node.js installed on your local machine to execute the scanner, and you need read access to your agent's configuration directories (e.g., `~/.claude` or `~/.codex`).

    New in LLM Ops