cogsec-hardening

claude-cogsec

English · 繁體中文

Claude Code Cognitive Security — a hardening pack for Claude Code, packaged as both a Claude Code plugin and a manual install script.

Implements the configuration prescribed by the cogsec / affaan-m guides:

• Longform Guide to Everything Claude Code — token optimization, memory persistence, evaluation, parallelization.
• Shorthand Guide to Everything Agentic Security — attack vectors, sandboxing, sanitization, CVEs, AgentShield.

Sources

This plugin codifies the configuration prescribed by:

• affaan-m/ECC — The Guides
• The Longform Guide to Everything Claude Code
• The Shorthand Guide to Everything Agentic Security

The name cogsec = cognitive security: defending the model's reasoning surface against prompt injection, memory poisoning, supply-chain payloads, and agent-identity abuse — distinct from classical infosec which defends the network/host surface.

Contents

.claude-plugin/plugin.json     # plugin manifest
hooks/
  hooks.json                   # PostToolUse observability binding
  log-tool-call.js             # JSONL logger -> ~/.claude/logs/tool-calls.jsonl
commands/
  cogsec-audit.md              # /cogsec-audit slash command
skills/
  cogsec-hardening/SKILL.md    # consolidated reference
rules/common/
  token-optimization.md
  memory-persistence.md
  evaluation.md
  parallelization.md
  agentic-security.md
templates/
  settings.deny-list.json
  settings.full-example.json
install.sh / install.ps1

Prerequisite — install ECC first

claude-cogsec builds on top of affaan-m/ECC (Everything Claude Code). Install ECC before claude-cogsec — it provides the rule pack layout, agent definitions, and the GateGuard hook that claude-cogsec extends.

# Add the ECC marketplace
/plugin marketplace add https://github.com/affaan-m/ECC

# Install the ECC plugin
/plugin install ecc@ecc

Then proceed to install claude-cogsec below.

Installation

Option A - plugin marketplace

claude /plugin marketplace add felimet/claude-cogsec
claude /plugin install cogsec-hardening@claude-cogsec

The plugin registers the PostToolUse hook, installs /cogsec-audit, exposes the cogsec-hardening skill. Rule files are copied via the install script (rules are not yet plugin-distributable in Claude Code).

Option B - manual install

./install.sh          # Linux / macOS / WSL / Git Bash

./install.ps1         # Native Windows PowerShell

Then merge templates/settings.deny-list.json into ~/.claude/settings.json and add the hooks.PostToolUse block shown in templates/settings.full-example.json.

What the plugin does

1. Logs every tool call to ~/.claude/logs/tool-calls.jsonl (timestamp, session id, tool, command, files touched).
2. Adds /cogsec-audit - one-shot hygiene scan.
3. Installs cogsec-hardening skill - consolidated reference.
4. Provides deny list baseline - blocks reads/writes to SSH/AWS/GPG/.env*, blocks curl|bash, wget|bash, raw ssh|scp|nc.

Manual settings.json edits

{
  "autoApproveTools": false,
  "hooks": { /* templates/settings.full-example.json */ },
  "permissions": {
    "deny": [ /* templates/settings.deny-list.json */ ]
  }
}

Verification

ls -la ~/.claude/logs/tool-calls.jsonl
tail -1 ~/.claude/logs/tool-calls.jsonl
ls ~/.claude/rules/common/
claude --version              # >= v2.0.65
claude /cogsec-audit --deep   # after plugin enabled

CVE patch status

CVE	CVSS	Required version
CVE-2025-59536	8.7	Claude Code >= v1.0.111
CVE-2026-21852	n/a	Claude Code >= v2.0.65

Layout rationale

• rules/ mirrors affaan-m/everything-claude-code rule pack layout, mergeable with existing ECC installs.
• hooks/hooks.json uses ${CLAUDE_PLUGIN_ROOT} so plugin path resolves regardless of install location.
• skills/ uses Anthropic skill frontmatter so it loads via standard Skill tool.

Fact-Forcing Gate

ECC ships a GateGuard hook (pre:edit-write:gateguard-fact-force) that intercepts the agent before any new file is created and forces it to answer four questions. claude-cogsec cooperates with this gate — the questions are reproduced here so you understand what they enforce.

1. Who will call this new file?

Name the file(s) and line(s) that will call this new file.

The agent must name the files and line numbers that will import, require, read, or reference the new file. Ensures the new file is not an orphan and is actually consumed somewhere.

2. Confirm no existing file serves the same purpose

Confirm no existing file serves the same purpose (use Glob).