Skip to content
Back to search

raptor-loop-hunt

RAPTOR autonomous looping multi-altitude security vulnerability hunt — Claude Code skill

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/dinosn--raptor-loop-hunt/badge.svg)](/listings/dinosn--raptor-loop-hunt)

HTML

<a href="/listings/dinosn--raptor-loop-hunt"><img src="/api/listings/dinosn--raptor-loop-hunt/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

**As a plugin — no clone needed.** Add the marketplace and install; Claude Code fetches it for you:

/plugin marketplace add dinosn/raptor-loop-hunt
/plugin install raptor-loop-hunt@raptor

Then `/reload-plugins` (or restart). The skill auto-triggers on audit requests, or invoke it directly.

**Or drop it in as a personal skill** — clone into your Claude Code skills directory and it is discovered automatically on the next session:

git clone https://github.com/dinosn/raptor-loop-hunt ~/.claude/skills/raptor-loop-hunt

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

.claude-plugin/marketplace.json · Manifest · 15 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. {
  2. "name": "raptor",
  3. "owner": {
  4. "name": "Nicolas Krassas"
  5. },
  6. "description": "RAPTOR security-research skills for Claude Code.",
  7. "plugins": [
  8. {
  9. "name": "raptor-loop-hunt",
  10. "source": "./",
  11. "description": "Autonomous looping multi-altitude vulnerability hunt (generate -> judge -> verify).",
  12. "version": "1.0.1"
  13. }
  14. ]
  15. }
README.md · README · 148 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. # raptor-loop-hunt
  2. An autonomous, looping, multi-altitude **security vulnerability hunt** for a codebase, packaged as a
  3. Claude Code **skill**. It replaces the model's default "single pass, summarize, stop" behaviour with an
  4. explicit search procedure: traverse every altitude, generate candidates, adversarially verify them
  5. *from raw source*, run isolated parallel reasoners, and keep a persistent ledger so each loop is
  6. net-new coverage instead of rediscovery. In practice it finds far more real bugs than a one-shot scan —
  7. the "Karpathy auto-research" methodology.
  8. It triggers whenever you point Claude at source code and want vulnerabilities found — "audit this",
  9. "find every bug", "security-review it", "find anything exploitable".
  10. ## Install
  11. **As a plugin — no clone needed.** Add the marketplace and install; Claude Code fetches it for you:
  12. ```
  13. /plugin marketplace add dinosn/raptor-loop-hunt
  14. /plugin install raptor-loop-hunt@raptor
  15. ```
  16. Then `/reload-plugins` (or restart). The skill auto-triggers on audit requests, or invoke it directly.
  17. **Or drop it in as a personal skill** — clone into your Claude Code skills directory and it is
  18. discovered automatically on the next session:
  19. ```bash
  20. git clone https://github.com/dinosn/raptor-loop-hunt ~/.claude/skills/raptor-loop-hunt
  21. ```
  22. ## Use
  23. Invoke it explicitly:
  24. ```
  25. /raptor-loop-hunt
  26. ```
  27. …or just describe an audit task ("audit ./src for security", "find every exploitable bug in this
  28. service") and the skill triggers on its own. Point it at a repo, a service, a module, or a directory.
  29. ## What it does
  30. - **Multi-altitude generate → judge → verify loop.** Candidates are generated at every altitude
  31. (whole-system down to a single function), then *adversarially* judged and verified from raw source —
  32. never from the generator's own summary.
  33. - **Isolated parallel reasoners** so independent findings don't contaminate one another.
  34. - **Monotonic-scrutiny knowledge base** (`raptor-loop-kb`) that can only ever *raise* scrutiny across
  35. runs, never lower it — each loop is net-new coverage rather than rediscovery.
  36. - **Disposition ledger** (`raptor-loop-ledger`) — an engagement-scoped state machine that certifies
  37. every disposition transition (confirm / reject / downgrade / sweep) against an evidence receipt, so a
  38. real bug can't be quietly rejected and a severe finding can't be laundered down to Low.
  39. - **Fresh-operator handoff gate** — records tacit workarounds and their resolutions, then independently
  40. replays an immutable PoC/lab package without author coaching. Technical validity and owner-delivery
  41. readiness remain separate outcomes.
  42. - **Execution-auth broker** (`raptor-loop-exec`) — a live-PoC command runs only under a typed,
  43. least-privilege capability grant.
  44. - **Eval trap battery** (`eval/`) — 12 deterministic and orchestrator-driven axes (recall,
  45. FP-resistance, false-rejection, counterfeit-evidence, coverage-honesty, severity-downgrade
  46. integrity, …).
  47. **The prompt is the program.** In an agentic system the model's "algorithm" is whatever you tell it to
  48. be; this skill specifies that algorithm. The quality comes from the structure, not from any single
  49. clever instruction.
  50. ## Layout
  51. | Path | Purpose |
  52. |------|---------|
  53. | `SKILL.md` | The skill itself — methodology, altitudes, generate→judge→verify loop, severity rubric. |
  54. | `references/vuln-class-discovery.md` | Leaf-level per-class search procedure (source → sink → oracle → variants). |
  55. | `references/kb-schema.md` | Schema + storage for the monotonic-scrutiny cross-run knowledge base. |
  56. | `references/ledger-schema.md` | Schema + the disposition-time transition gate (candidate ledger + evidence receipts). |
  57. | `references/handoff-reproducibility.md` | Fresh-operator replay, intervention receipts, and owner-delivery readiness. |
  58. | `scripts/raptor-loop-kb` | Deterministic cross-run KB helper (can only ever raise hunt scrutiny). |
  59. | `scripts/raptor-loop-ledger` | Candidate state machine — certifies each disposition against an evidence receipt. |
  60. | `scripts/raptor-loop-exec` | Execution-auth broker — least-privilege sandbox plan for live PoCs. |
  61. | `scripts/test-*` | Self-contained tests (no network, no external imports). |
  62. | `eval/` | The 12-axis trap battery. |
  63. The scripts are pure-stdlib Python and self-contained; the execution broker will optionally use a host
  64. RAPTOR checkout's sandbox when one is present (`RAPTOR_DIR`), and refuses to run live commands
  65. otherwise.
  66. ## Track record
  67. **So far: 200+ verified vulnerabilities across 40+ real-world codebases (Dec 2025 – Jul 2026).**
  68. Run against C network daemons, message brokers, routing stacks, web apps, language runtimes, and media
  69. libraries, the skill has surfaced original, verified vulnerabilities across most of them. Findings are
  70. listed by title and class only; **exploit mechanics and PoCs are withheld**, and outcomes are stated at
  71. their true strength (a proven DoS is not called an RCE; unweaponized primitives are labelled as such).
  72. All have been reported to or communicated with the respective vendor / maintainer.
  73. ### Original findings
  74. | Target | Highest verified outcome | Class | Status |
  75. |--------|--------------------------|-------|--------|
  76. | memcached | pre-auth heap buffer overflow, default config, client-reachable | memory-corruption | Fixed upstream, finder credited |
  77. | Immich | unauth restore endpoint → OS-command-exec RCE (container-root) + broken-authorization / IDOR cluster | authz / command-exec | Disclosed |
  78. | RustDesk (client + server) | malicious-host file-transfer path-traversal → RCE on controller; secure-channel / persistent-ID-hijack primitives | path-traversal / crypto | Reported |
  79. | Samba 4.25 (master) | 2× pre-auth memory-corruption crash (DNS TSIG underflow; LDAP type-confusion) — DoS-grade | memory-corruption | Reported |
  80. | Apache Kafka (trunk / 4.0.0) | pre-auth secret / file disclosure + static-init RCE-class primitive | authz / code-load | Reported |
  81. | FRR (master) | pre-auth EIGRP authentication bypass + ASAN-proven FPM stack overflow | auth-bypass / memory-corruption | Reported |
  82. | Redis (unstable / 8.8.0) | pre-auth AUTH/ACL bypass + RCE-grade primitives (not weaponized to full shell) | auth-bypass / memory | Reported |
  83. | Linux XDP zero-copy | OOB over-copy in `xdp_build_skb_from_zc`; root RCE demonstrated on an instrumented rig | memory-corruption | Reported |
  84. | Apache ActiveMQ 5.19.6 | authenticated RCE + multiple authz-bypass / IDOR primitives (incl. a pre-auth connection-kill) | authz / deser | Reported |
  85. | LimeSurvey 7.0.1 | unauth time-based SQL injection + zip-slip → webroot RCE | injection / path-traversal | Reported |
  86. | Apache Axis 1.x / 2 | unauth deserialization RCE (lab-proven) | deserialization | Reported |
  87. | Apache Tomcat 12 (dev) | unauth blind SSRF / DNS-exfil (Medium) | SSRF | Reported |
  88. | Quagga 1.2.3 / 1.2.4 bgpd | pre-auth stack OOB write + use-after-free (DoS-confirmed) | memory-corruption | Reported |
  89. | llama.cpp | pre-auth process-crash DoS (RCE escalation hypothesized, not proven) | memory-safety | Reported |
  90. | Fastjson 1.2.83 | RCE (lab-proven) | deserialization | Public |
  91. | Pillow 12.3.0 | blind-calibration run: rediscovered a real heap-OOB write + 4 net-new ASan-confirmed bugs | memory-corruption | Reported |
  92. **Adversarial rigor, not just recall.** The judge/verify half of the loop is built to *refute*, so it
  93. also kills bad findings: it disproved a fabricated pre-auth RCE exploit package live, and corrected an
  94. over-stated public "ASLR bypass" down to what actually reproduces.
  95. ### Public CVEs reproduced / validated
  96. Independent labs, PoCs, and patch analysis — validations of already-public issues, not original
  97. discoveries:
  98. | ID | Component | Class | Notes |
  99. |----|-----------|-------|-------|
  100. | CVE-2026-42588 / -42253 | Apache ActiveMQ | Jolokia → XBean RCE; header-injection / XSS | reproduced end-to-end |
  101. | CVE-2026-15583 | Grafana MCP (`mcp-grafana` ≤ 0.17.1) | confused-deputy SA-token exfil | reproduced; fixed 0.17.2 |
  102. | CVE-2026-20230 | Cisco CUCM 14 / 15 | pre-auth SSRF → Axis 1.4 → JSP webshell | reproduced (in-the-wild chain) |
  103. | CVE-2026-4282 | Keycloak < 26.5.7 | PAR `request_uri` → forged OAuth code → admin token | lab A/B validated |
  104. | CVE-2024-21182 | Oracle WebLogic 12.2.1.4 / 14.1.1 | T3 / IIOP JNDI-injection RCE | reproduced; public lab repo |
  105. | CVE-2025-13673 | Tutor LMS (WordPress) ≤ 3.9.7 | unauth SQLi via `coupon_code` | UNION exploit + Docker lab |
  106. | MDEV-39493 (ZDI-CAN-30333/4) | MariaDB ≤ 11.4.10 | FILE-priv bypass via CTE `INTO OUTFILE` | reproduced; fixed 11.4.11 |
  107. | CVE-2026-6479 | PostgreSQL | pre-auth SSL/GSS unbounded recursion (DoS) | reproduced (needs `-O0` build) |
  108. | ZDI-CAN-30607 | dnsmasq | `find_soa()` out-of-bounds read | ASan-confirmed + 22-site variant sweep |
  109. | CVE-2026-4176 | Perl `Compress::Raw::Zlib` ≤ 2.219 | `crc32_combine` DoS | reproduced; reachability assessed low |
  110. | CVE-2026-27654 | nginx (WebDAV COPY/MOVE) | heap overflow | reproduced; fixed 1.29.7 |
  111. | CVE-2026-42945 (Rift) | nginx | primitive-recovery chain | partial; corrected the public ASLR-bypass framing |
  112. | CVE-2026-35273 | Oracle PeopleSoft (PSEMHUB) | unauth SSRF → XXE → webshell | analysis-only (JAR patch-diff) |
  113. ## Scope
  114. For defensive security research, education, and authorized penetration testing only.
  115. ---
  116. > Oh — and RAPTOR is required. Obviously. It's right there in the name. 🦖
  117. > Go feed the bird → [github.com/gadievron/raptor](https://github.com/gadievron/raptor)

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.