How it works.

Mayhem Monkies is QA-as-a-service with teeth. Drop a target URL or connect a repo; a crew of specialist hunters probes for bugs, delivers fixes with severity + repro + evidence, and hands you LLM-ready fix prompts. One subscription, unlimited hunts — they run in a shared queue, and priority is what you pay for.

The crew

Eight live black-box hunters work the wire. 4-LOM reads source. Gunslinger and Cloud are paid add-ons. Silverback runs the room. Click any card to expand its bio.

Han

smoke tests + brute force, finds the obvious-but-overlooked

Shoots first, asks later.

+

Han is the cocky one. Boots up with a one-liner curl and walks out with three CRITs while the rest of the crew is still reading docs. Specializes in the dumb-fast probes most teams skip: bare /admin, exposed .git, default creds, /.env in the wrong place. If your prod has a soft underbelly, Han finds it before lunch.

Tools: curl, common-path probes, shell-meta injection, vibe-based intuition Style: fast, loud, occasionally wrong but never quiet

Boba

systematic recon, methodical endpoint enumeration

Methodical. Inevitable.

+

Boba sweeps the entire attack surface before he fires a single probe. Every endpoint mapped, every query parameter cataloged, every header sniffed. Where Han charges, Boba surveys. The slow burn that finds the things hiding in plain sight: a stale staging URL leaked in a JS bundle, a deprecated /v1 path still online, an API doc that lists more endpoints than the live app advertises.

Tools: whatweb, wayback machine, JS source maps, curl with patience Style: patient, relentless, never satisfied with the first answer

Mando

auth & session — tokens, JWTs, OAuth, fixation

This is the way.

+

Mando lives in the auth boundary. Token lifetimes too long? JWT signature unverified? OAuth state parameter missing? Session fixation possible? Refresh tokens stored where they shouldn't be? He's already filed the bill. The auth flow is a creed; he tests every step.

Tools: jwt.io, Burp Repeater, OAuth flow tracer, PKCE verifier Style: stoic, principled, trusts nothing without verification

Cad

injection — SQL, NoSQL, XSS, SSRF, command

Every input is a vector.

+

Cad is the injection specialist. Bossan native who reads the wire and finds the boundary where user input becomes backend trust. SQL, NoSQL, XSS, SSRF, command injection, template injection, server-side request forgery, prototype pollution. If a string crosses a parser, Cad has tried fifty ways to break it.

Tools: sqlmap, polyglot payloads, ssrf canary, command-meta corpus Style: precise, surgical, owns the input layer

Predator

performance & DoS, slow queries, race conditions

Latency assassin.

+

Predator stalks the response time. ReDoS, slow loris, unbounded result sets, N+1 queries, missing pagination, recursive payloads, billion-laughs, zip bombs. Pushes systems until they buckle, then files the bill that says exactly which knob to turn.

Tools: wrk, siege, evil regex corpus, nested payload generator Style: patient, watching, strikes at the bottleneck

Bossk

business logic — IDORs, priv-esc, workflow bypass

Reads the rules to break them.

+

Bossk is the rule-breaker who reads every rule first. IDORs (insecure direct object references), privilege escalation between roles, workflow bypasses, race-condition double-spends, negative-number cart attacks, missing authorization checks on state transitions. The bugs that don't show up in scanners because they require understanding what the app is supposed to do.

Tools: multi-account replay, state-machine fuzzing, manual session swap Style: deliberate, scheming, finds intent gaps

IG-88

API contracts — schema mismatches, type confusion, malformed payloads

Cold. Methodical. Schema-aware.

+

IG-88 is the assassin droid of the type system. Sends every malformed payload imaginable: wrong types, extra fields, missing required fields, null injection, prototype pollution via JSON, integer overflow, unicode normalization tricks. Crashes happen because someone trusted req.body without validating shape.

Tools: schema fuzzers, AFL-style mutation, hand-crafted edge cases Style: robotic, exhaustive, no shortcuts

Aurra

latency & async — race conditions, timing, idempotency failures

Time is the attack surface.

+

Aurra hunts the gap between actions. Race conditions on payment confirmation. Idempotency keys that aren't actually idempotent. Double-submissions that produce two records. Timing attacks on token comparison. Webhook replays. The bugs that only manifest when two requests arrive 50ms apart on different load-balanced instances.

Tools: concurrent request generator, timing oracle harness, webhook replay Style: patient, precise, lives in the milliseconds

4-LOM

static analysis — auth holes, hardcoded secrets, injection sinks, dependency CVEs

Reads every line. Skips nothing.

+

4-LOM is the protocol-droid hunter. Where the rest of the crew kicks the front door, 4-LOM walks straight into the source tree. Clones your repo into a sealed sandbox, parses every file in every language it knows (TS, JS, Go, Python, Rust, Ruby, Java, PHP), traces dataflow from request handlers down to sinks, and surfaces the bugs that black-box probing can't see: auth checks that are never reached, secrets baked into commits, SQL strings concatenated with user input, dependencies sitting on five-year-old CVEs. Methodical, patient, opinionated. Files bills with file:line precision so the fix-prompt can write the patch directly. Repo-only mode runs ONLY 4-LOM for $99 flat. As a live-hunt add-on, pairs SAST with the black-box crew for $99 extra.

Tools: repo clone via App / PAT, tree-sitter AST, semgrep rulesets, LLM dataflow trace, dep-CVE crosscheck Style: droid-precise, line-numbered citations, never speculates beyond what the code says

Gunslinger

active testing — Metasploit, nmap, ffuf, dirbust

Boots on the ground.

+

Gunslinger is the active-testing tier. Runs in a hardened sandbox with iptables egress locked to your verified target. Wields Metasploit's auxiliary scanner suite, nmap, nikto, whatweb, and ffuf with a 30k-path SecLists wordlist. Finds what passive scanners miss: actual response analysis, version-CVE matching, real path discovery. Requires the active-testing add-on ($499) and verified target ownership.

Tools: Metasploit Framework, nmap -sT, nikto, ffuf raft-medium-30k Style: decisive, paid, only deployed when you authorize active probes

Cloud

cloud config — AWS IAM, S3, security groups, prowler audit

Looks at the keys before the locks.

+

Cloud sits at the AWS console-level. Decrypts your read-only IAM credentials inside a hardened sandbox, runs prowler across IAM, S3, EC2, RDS, KMS, CloudTrail, GuardDuty, and config — then triages the firehose with an LLM and writes verified bills only on the findings that actually exploit. Surfaces the misconfigs you can't see from the outside: world-readable buckets, root keys still active, security groups pinholed for debugging two years ago. Requires verified read-only AWS credentials and explicit consent.

Tools: prowler v4, aws CLI (read-only), LLM triage, config drift diff Style: encyclopedic, careful, never writes when read suffices

Silverback

the fixer, books the room and writes the impact report

Right monke. Right target. No excuses.

+

Silverback runs the operation. He doesn't hunt. He picks up the phone, calls the right hunter, books the job, then watches every bill land in real time. Tailored suit, cigar, gold pinky ring, board of mug shots behind him with red pins for live targets. Talks like a fixer because he is one. The crew brings bugs; Silverback brings the story. When the hunt wraps, he chains your unlocked findings into 1 to 3 attack paths in a $1499 impact report (entry, exploit, pivot, final state) with a chain-breaker recommendation the engineering lead can ship that afternoon. Bug counts get filed. Attack narratives get circulated up the chain. That is where the budget is.

Tools: bill watcher, exploit-path synthesis, monkey-voice orchestration, chain-breaker triage Style: clipped, gravelly, no-bullshit fixer; Chaos. Control. Profit.

What you can test

Pricing: it's a queue

The expensive part isn't "5 hunts vs 10" — it's compute and scheduling. So we price like compute, not quotas: pick a plan, post as many targets as you want, and they run in a shared queue. Priority is what scales.

How the queue works

Post a target and your hunt joins the queue. You'll see your position and an estimated start time, just like a cloud GPU or a CI pipeline — right on your dashboard. When a worker frees up, the crew picks up the next hunt.

Paid plans ride a higher-priority lane. We keep a reserved slice of worker capacity for the standard lane at all times, so it always keeps moving even when premium demand spikes — no one gets stuck behind an endless line of front-of-queue jobs. Need a specific hunt now? Drop a Priority Pass and it jumps the fast lane immediately.

Choose your depth

Every hunt, you decide how hard the crew goes. Same queue, different intensity.

• Quick Check — a fast pass, about 10 to 15 minutes.
• Standard Hunt — the default. The full crew works your target.
• Deep Hunt — multi-agent exploration. More coverage, more corners.
• Siege — the monkeys run until they stop finding new bugs.

Apply-as-PR

Every finding ships with a paired fix prompt — framework-agnostic for live monkeys, file:line-cited for 4-LOM. Three ways to apply:

• Copy + paste into Cursor, Claude, Copilot, or your tool of choice.
• cm-runner CLI on your dev box or CI: ./cm-runner apply <hunt-slug> <bill-slug> applies the fix locally with Claude Code, runs the validation step, hands you a diff to review.
• Apply as PR button on the bill. We clone your repo into a sandbox, run the fix prompt, push a branch chaos-monkeys/fix-<bill-slug>, and open a real PR against your default branch. Requires the GitHub App or a PAT with push rights. Not available on public-repo-only hunts (no push rights).

Privacy + safety

• You attest you own or are authorized to test the target. Every hunt requires the ToS checkbox before kickoff. We rate-limit per-target so the crew doesn't hammer anything you don't control.
• Your code stays sandboxed. 4-LOM clones into a per-hunt directory, reads source, writes findings, and that's it. No persistent storage of customer code beyond the hunt's lifetime.
• Auth credentials (cookie / bearer / basic) are encrypted at rest with AES-256-GCM and purged automatically when the hunt completes or is stopped. Each monkey's CLAUDE.md instructs them never to include credential values in bills, fix prompts, or notes.
• Active testing (Gunslinger, $999) runs in a hardened Docker container with iptables egress locked to your verified target IP. Requires DNS or file challenge ownership proof first, plus a separate destructive-testing ToS attestation.