top of page

We Ran Claude BugHunter: 83 Red Team Skills for Your AI Agent

3 minutes ago
6 min read

A bug bounty toolkit that turns your coding agent into a red team assistant is a big promise. If you run an MSP, you can already hear the pitch: point it at a client's external footprint and let it hunt. Before anybody on your team does that, here is what we saw when we actually ran it.

What Did We Actually Do?

On Monday, Nov. 16, 2026, we installed Claude BugHunter in a locked down sandbox at commit 2cfe533 (package version 2.1.0). We ran the installer two ways, uninstalled it, ran the cbh command line tool, ran the hunting engine in its mock mode, ran its scope self test, and scaffolded a practice engagement folder.

We did not run a live hunt against anything real. We did not install it through the plugin marketplace, connect Burp Suite, run the community installer, or test Windows. We also did not read through all 83 skills one by one. Everything below comes from our own run logs or from the public repo at that commit.

What Is Claude BugHunter?

Claude BugHunter is a bundle of skills for Claude Code aimed at bug bounty and external red team work. At the commit we tested, it ships 83 skills, 15 slash commands, a terminal tool called cbh, and an engagement engine. We counted the skill folders ourselves, and our install log agreed: "83 skills installed."

The skills load automatically by topic. You describe what you are testing in plain English and the relevant skill kicks in, with no invocation by name. The README also says the same skills install into OpenCode, OpenAI Codex CLI, Hermes Agent, and Google AntiGravity.

It had 6,998 GitHub stars as of Nov. 16, 2026, when we tested commit 2cfe533.

What Was the Good News?

Credit where it is due:

  • Open licenses. The code is MIT licensed and the skill content is CC BY 4.0. The credits file names the outside project it vendors and that project's MIT license.

  • No calls home. In the code we read, the network calls go to a public certificate transparency site (crt.sh) and the CISA Known Exploited Vulnerabilities feed. We found no telemetry.

  • A clear authorization stance. The README says the skills are for assets you own or have written authorization to assess. It points offensive security folks to Anthropic's Cyber Verification Program and says plainly, "Do not reword an offensive engagement to look defensive to get past a classifier."

  • The engine's scope check works. The engine's scope list is default deny. In our run, its self test passed and it rejected our made up host as out of scope.

  • Hunting is opt in. A default engine run stops at a map of the target. Agents only get spent when you pass --hunt, and intrusive actions stay off unless you allow them.

  • A clean uninstall, mostly. The installer writes a manifest. In our run, uninstall "removed 99 item(s)" and took its shell line back out.

  • Honest labeling. The README admits its demo transcript is illustrative.

We found nothing malicious. No hidden characters, no prompt injection aimed at your agent, no piping a download into a shell, no credential harvesting.

What Is the Fine Print?

1. The recon command does not check scope. The README describes a workflow "with scope enforced in code." That is true for the engine. It is not true for cbh recon. We ran cbh recon against a made up domain with no scope file. It tried crt.sh, then sent a web request to that host, the same host the engine's scope check had just rejected. Our stand in server logged the request. One typo in a target name, and you are knocking on someone else's server.

Fine print 2: what it says is read only agents, what it runs is permission prompts off, with permission mode bypassPermissions, curl and python3 allowed

2. "Read only" is a sentence in the prompt. The engine README says its agents run "read-only by default." In the code, each agent is launched with --permission-mode bypassPermissions, with curl and python3 on the allowed tools list. We captured exactly that in our run. The read only part is an instruction written into the agent's prompt. Anthropic's own Claude Code docs say "Permission rules are enforced by Claude Code, not by the model," and that bypassPermissions should only be used "in isolated environments like containers or VMs."

Fine print 3 and 4: findings go into one shared memory file even with the memory option off, and the Burp option edits Codex and OpenCode settings that uninstall left behind in our run

3. Every confirmed finding lands in one shared file. When the engine confirms a finding, it writes it to ~/.claude/bughunter/memory/findings.jsonl. That file is cross target by design, and the write happens even when you do not turn on the memory option. In our mock run, without that option, the file showed up with the test finding in it. For an MSP, that means Client A's findings and Client B's findings sit in the same plain text file unless you change where it goes. You can, with BUGHUNTER_MEMORY_DIR.

4. The installer reaches past Claude. By default, the installer adds a line to your shell startup file (--no-shell skips it). The --burp-mcp option writes Burp entries into your Codex and OpenCode settings. In our run, uninstall removed everything in its manifest, but those Codex and OpenCode entries stayed behind.

One nit. Phrases like "battle-tested" and the "681 disclosed-report patterns" count are the project's own claims. We did not test them.

What Should You Do If You Try It?

  1. Use a dedicated testing machine or VM. Never your daily laptop or a client device.

  2. Work from a written scope file, and double check every domain you type. The recon command will not catch a typo for you.

  3. Never run agents with permissions bypassed outside an isolated box.

  4. Give each client its own memory folder, and wipe it when the job ends. Set BUGHUNTER_MEMORY_DIR per engagement.

These four rules are Chibitek guidance, based on our sandbox run and the project's own code and docs.

Key Points of Claude BugHunter

  • Claude BugHunter is a bug bounty and red team skill bundle for Claude Code: 83 skills, 15 commands, a cbh CLI, and an engagement engine at commit 2cfe533.

  • Chibitek installed and ran it in a sandbox on Nov. 16, 2026. Verdict: PASS WITH NITS. Nothing malicious found.

  • The engine's scope list is default deny and works. The cbh recon command has no scope check.

  • The engine's "read only" agents run with Claude Code's permission prompts bypassed.

  • Confirmed findings from every target go into one shared file by default.

  • The Burp option edits Codex and OpenCode settings that uninstall left behind in our run.

Frequently Asked Questions

Did you run a real hunt? No. We ran the engine in mock mode, with a stand in for Claude, inside a sandbox. No live AI hunt, and no real targets.

Is Claude BugHunter open source? Yes. The code is MIT licensed and the skill content is CC BY 4.0.

Does it send data anywhere? In the code we read, outbound calls go to crt.sh for certificate lookups and to the CISA vulnerability feed, plus whatever targets you point it at. We found no telemetry and no calls home.

What does bypassPermissions mean? It is a Claude Code permission mode that skips the approval prompts. Anthropic's docs say to use it only in isolated environments like containers or VMs. Anything the agent is told not to do after that is up to the model, not the permission system.

Let's Review Before You Install

At Chibitek, we look closely at AI tools before they get anywhere near a client, with no geek-speak. An AI agent that can hunt for bugs is a powerful thing to hand your team. Give it its own machine, a written scope, and its own memory per client. Contact us at hello@chibitek.com to talk through how you test tools like this, and where they should run.

Test machines only. Written scope. Isolated agents.

We make IT effortless so you can create, disrupt, and grow!

Follow Chibitek on YouTube at @chibitekmsp, and follow Chibitek on LinkedIn.

Sources

Every repo claim was checked against the file at the tested commit (2cfe533df8efa2ced87c91299e1f62b88cb26108, committed Nov. 15, 2026, 6:59 AM ET), fetched from GitHub on Nov. 18, 2026 and byte compared with our test copy, and then against Chibitek's own sandbox run logs from Nov. 16, 2026.

Comments


bottom of page