0xSteph/pentest-ai-agents

What it is

pentest-ai-agents (0xSteph) is a collection of 50 Claude Code subagents that turn Claude into an offensive-security assistant. Each agent carries domain knowledge for a specific area — recon, web, Active Directory, cloud, mobile, wireless, social engineering, payload crafting, reverse engineering, exploit chaining, detection, forensics. It ships as pure agent files (markdown/frontmatter): "no servers, no Python deps." You describe a task in Claude Code and it routes to the right specialist.

How it works

Agents are split into two tiers. Tier 1 (advisory) agents are routable from any task. Tier 2 (execution-capable) agents can run Bash, and each must carry a mandatory _scope-guard.md scope-enforcement block; they live in a separate "offensive operations cluster" and require a declared scope. A hardened, SHA-pinned CI validator checks every agent's frontmatter, enforces the scope-guard on all Tier 2 agents, and smoke-tests the installer. Slash commands include /recommend "task" (routes to an agent + concrete commands) and /agents-for <tag> (filters by domain). A findings DB tracks vulns (v2 adds a tool_used column and CVE/tool indexes).

The offensive coverage is real and specific: c2-operator (Sliver/Mythic/Havoc/Cobalt Strike), container-breakout (Docker/K8s escape, runc/cri-o CVEs, kubelet, RBAC abuse), data-exfiltrator, evasion-specialist, persistence-planner, lateral-movement, opsec-anonymizer (burner infra, source-IP design), llm-redteam (OWASP LLM Top 10, prompt injection, RAG poisoning, MCP abuse), and phishing-operator (Evilginx, GoPhish, dnstwist).

What to do

Primary install is a piped shell script that clones to temp, copies agents to ~/.claude/agents/, and exits (idempotent):

curl -fsSL https://raw.githubusercontent.com/0xSteph/pentest-ai-agents/main/install.sh | bash

Or install as a Claude Code plugin:

/plugin marketplace add 0xSteph/pentest-ai-agents
/plugin install pentest-ai-agents@pentest-ai-agents

Optional helpers: install.sh --tools pulls underlying CLIs via apt/brew/pacman + pipx/go/cargo; db/doctor.sh audits which tools are installed per agent; --uninstall removes everything.

Caveats

• The safety model is guardrails, not sandboxing. A _scope-guard.md hard-refusal list covers DoS, mass scanning, unattended worms, false-flag operations, and safety-of-life systems — but enforcement is prompt-level, and the page itself notes a scope-guard was recently missing from the Bash-capable cicd-redteam agent until CI caught it. Treat scope enforcement as best-effort.
• Evaluate the curl | bash install path against your own supply-chain policy; it fetches and executes code from main at install time.
• Capabilities like data-exfiltrator, c2-operator, and evasion-specialist are genuinely offensive — this belongs only in authorized, scoped engagements.