Claude Code gets 35 offensive security hats

What it does

This is essentially a curated library of 35 prompt-based subagents that drop into Claude Code’s agent directory. Describe a task—planning an AD pentest, reversing firmware, or auditing STIGs—and Claude routes to the relevant specialist such as ad-attacker or stig-analyst. The agents cover the full kill chain from reconnaissance to reporting, but the actual exploitation is still performed by external tools you bring yourself.

The interesting bit

The authors treat Claude less like a chatbot and more like a shell for spawning domain experts. A built-in scope guard hard-refuses tasks like DoS or mass scanning, and a two-tier system keeps advisory agents open while execution-capable ones require a declared engagement scope. There is even an llm-redteam agent for testing other LLMs, which feels appropriately meta.

Key highlights

• 35 agents spanning web, Active Directory, cloud, mobile, wireless, social engineering, C2 operations, and reverse engineering
• Ships with a SQLite findings database and a doctor script to audit which of the 80+ underlying CLI tools are present on your machine
• Tier 1 agents give advice; Tier 2 agents can execute tools directly, but only after scope is declared
• Recent additions include container-breakout, c2-operator, opsec-anonymizer, and llm-redteam
• No persistent server or Python dependencies—just agent files copied into Claude Code’s config directory

Caveats

• This is fundamentally a prompt library and shell glue; the heavy lifting is done by external tools like BloodHound, Nuclei, or Cobalt Strike that you must install separately
• Capabilities depend heavily on your local tooling; the included doctor script exists mainly to show you what is missing
• The quick-start install pulls from the live repository, which may give pause to the paranoid

Verdict

Worth a look if you are a pentester or red-teamer already using Claude Code and want structured domain expertise without context-switching. Skip it if you are not doing authorized offensive security; the scope guard is explicit, but the tooling is clearly aimed at professionals.