Who audits the code that AI writes? As generation gets faster, review becomes the bottleneck. As of November 17, 2026, three repositories on GitHub offer three sharply different answers, and this comparison puts them on one table. One hands judgment to isolated verification agents that never see the finding before it reaches them. One pins the parts that must not go wrong to deterministic engineering. One does not audit code at all, turning a coding assistant into a red-team operator for the external attack surface. So this is a review of Three Tools and Three Licenses: Audit, Review, and Bug Hunting. We check only what the README, LICENSE, NOTICE, and GitHub API fields can support. There is no buying advice, no benchmark ranking, and no sentence added on any project's behalf that its own documents do not say.
At a glance
The three repositories are cloudflare/security-audit-skill, alibaba/open-code-review, and elementalsouls/Claude-BugHunter. Stars, forks, and issues are written only from the GitHub API snapshot as of November 17, 2026, never blended with the monthly increments in third-party trend badges.
| Dimension | Cloudflare security-audit-skill | Alibaba open-code-review | BugHunter |
|---|---|---|---|
| License | MIT, copyright line Copyright (c) 2025-2026 Cloudflare, Inc. | Apache-2.0, appendix credit 2026 alibaba/open-code-review Contributors | MIT code plus CC BY 4.0 content, dual license; API reports MIT only |
| Stars (Nov 17, 2026 snapshot) | 73,930 | 52,922 | 7,453 |
| Language and form | JavaScript, general coding-agent skill | Go, AI code review CLI | Python, Claude Code skill bundle |
| Created, pushed | 2026-06-18, 2026-09-14 | 2026-05-18, 2026-11-16 | 2026-05-05, 2026-11-16 |
| Open issues | 129 | 335 | 4 |
| One-line position | Turns your agent into a security auditor | Deterministic pipelines plus LLM agent hybrid | Bug hunting and external red-team work |
The license row is the one to stare at. Cloudflare is a one-page MIT whose copyright line uses the 2025-2026 range rather than a single year. Alibaba is a standard Apache-2.0 text, but its copyright holder sits in the LICENSE appendix, and the README carries a differently worded credit line; both are recorded separately here rather than merged. BugHunter looks like plain MIT to the GitHub API, yet the repository ships two license files, code under MIT and documentation and skill content under CC BY 4.0, so the API reports only half the picture. The Three Licenses in the title are not three phrasings of the same thing; they are three different authorization structures.
Three licenses, from one page to a dual grant
Cloudflare's LICENSE is standard MIT, copyright line verbatim Copyright (c) 2025-2026 Cloudflare, Inc., with a single condition: copies and substantial portions must retain the copyright and permission notices. The year range matters; quoting it as 2026 alone is wrong.
Alibaba's LICENSE is Apache-2.0, with an explicit patent grant and a NOTICE clause, but the repository root carries no NOTICE file. A clause in the text and no file in the project are both stated as they stand. The copyright also appears in two places: the LICENSE appendix names alibaba/open-code-review Contributors, the README names Alibaba. One project, two wordings, kept apart.
BugHunter's dual license is the least intuitive cell in this comparison. Its MIT covers source files such as *.py and *.sh; CC BY 4.0 covers *.md documentation, wordlists, regex catalogs, and scoring rubrics. The README says the content may be reused freely, including commercially, provided you credit Sachin Sharma / Claude-BugHunter and link back to the repository. The NOTICE adds that the repository does not redistribute upstream community skills directly; users install them from original sources through a bootstrap script. The GitHub API license field returns only MIT, and treating BugHunter as pure MIT free for uncredited commercial use would run straight into the CC BY 4.0 attribution duty. The three licenses tighten from permissive to conditional, which maps exactly onto the question of who audits and who bears responsibility.
Three review philosophies: who holds judgment
Cloudflare gives judgment to isolated verification agents. The README says the skill turns your agent into a security auditor, and the audit runs in six phases: reconnaissance, coverage-led hunting, candidate validation, structured output, independent record verification, and target-neutral reporting. Two design choices are what actually make independent verification real: fresh agents verify final source claims, with material replacements getting another independent verifier; and the agent that checks a finding is never the agent that found it. Verdicts come in three distinct grades: confirmed needs a complete source trace, needs_validation keeps an unresolved fact with no severity, and rejected records a disproved candidate. The six phases and the independent verification are one package, Cloudflare's alone, not the other two projects'. The project argues this conservatism buys trust, preferring to leave a conclusion at needs_validation rather than confirm a boundary it cannot support. Its structured output is equally deliberate: confirmed, needs_validation, and rejected records are written to findings.json and validated against report-schema.json. The validate-findings.cjs script runs in Phase 4 and again after every Phase 5 replacement, while validate-coverage-ledger.cjs runs from Phase 1 through Phase 5 after ledger creation and each later update; both validators are zero-dependency. Triggering is automatic by request semantics, so a security audit, finding vulnerabilities, or pen-testing the code all invoke it, and in full audit mode an unspecified output directory defaults to a fixed path. The stated design principles are worth quoting: only confirm established boundary failures, base severity on impact rather than deviation from a checklist, treat defense-in-depth gaps as not vulnerabilities, and expect multiple runs to be additive.
Alibaba open-code-review pins what must not fail to deterministic engineering. Its core design is Deterministic Engineering times Agent Hybrid, in other words deterministic pipelines paired with an LLM agent: file selection, smart bundling, fine-grained rule matching, and separate positioning and reflection modules are guaranteed by engineering logic, while the agent handles dynamic decisions such as scenario-tuned prompts and toolsets. It reads Git diffs and generates precise line-level comments, with built-in multi-language rules for NPE, thread safety, XSS, and SQL injection, compatible with OpenAI and Anthropic. The point is not a multi-phase audit but reliability pressed into hard constraints, leaving the language model only what it does best. The project also states its origin: an internal AI code review assistant at Alibaba that served tens of thousands of developers before being incubated as open source. The pain points it targets are written into the README: incomplete coverage, position drift, and unstable quality, with the root cause pinned on a purely language-driven architecture that lacks hard constraints on the review process. Its product form is a CLI plus host plugins across Windows, macOS, and Linux, covering Claude Code, Codex, Cursor, and Kimi Code, with OpenCode, QCA Forward, and skill-compatible agents documented as well, an MCP server, and CI integrations for GitHub Actions, GitLab CI, GitFlic CI, and Gerrit. Beyond reading diffs, an ocr scan mode reviews entire files for auditing unfamiliar codebases or directories with no meaningful diff.
BugHunter is not in the business of auditing AI-written code at all. It is a bug-hunting, external red-team bundle covering the external attack surface, meaning anything reachable from the internet without first compromising an internal endpoint. The README header states 83 skills and 15 slash commands, with knowledge drawn from 681 disclosed reports of which 433 are individually cited and auditable, organized into 24 core vulnerability classes. Its triage-validation and 7-Question Gate are submission gates, not Cloudflare-style independent verification; the wording differs and the two must not be confused. The authorization posture is hard-edged too: the skills are for assets you own or have written permission to assess, explicitly excluding weaponizing zero-days against unauthorized targets, post-exploitation tooling, and mass-targeting infrastructure. It even carries a runtime note that real-time network safeguards may block offensive tooling development and that authorized work can apply for separate verification. Structurally it has four layers: a thinking layer for methodology and red-team mindset along a non-linear workflow; a web-app layer with 58 hunt skills drawn from disclosed reports; a perimeter layer covering M365 and Entra, Okta, vCenter, SSL-VPN appliances, SharePoint, and cloud IAM with 2024 to 2026 CVE chains and post-credential escalation; and a shipping layer for triage-validation, reporting, and evidence hygiene, including VRT-aware severity, out-of-scope rebuttals, and PII redaction. The skill inventory reconciles: 58 web app vulnerability classes, 10 enterprise platform attack, 6 reporting and validation, 5 recon and OSINT, and 4 methodology and mindset. It also states its lineage, with a vendored foundation from another bounty project and a sister project for OSINT.
Installation, authentication, and namespaces are not interchangeable
Treating the three as swap-in plugins is another common misreading. Cloudflare is a general coding-agent skill loaded through the Skills CLI; its README requires no API key. It runs inside your own agent, needs a model with tool use and parallel sub-agents, and needs an OS-enforced, network-disabled sandbox for target-controlled builds. Alibaba is an AI code review CLI plus host plugins, an MCP server, and CI integrations; the default mode requires you to configure a model endpoint, Delegation Mode is what removes the OCR key requirement, and it explicitly does not consume a Claude Code Team or Pro subscription login. BugHunter is a Claude Code skill bundle installed through the plugin command; its core skills and CLI run inside the Claude Code subscription with no separate key, and only the optional skill regenerator needs an Anthropic API key. The three have incompatible install entry points, namespaces, and update mechanisms, so it is wrong to write that all three need an API key, all three work as Claude Code skills, or all three are MIT.
The capability domains are misaligned as well. Cloudflare and Alibaba both face code review; BugHunter faces the external attack surface, listing internal Active Directory, C2, post-exploitation, lateral movement, and evasion as out of scope. The three do not even share a command set. Cloudflare loads on trigger phrases, Alibaba ships a CLI and host plugins, and BugHunter's slash commands and /hunt engine stay Claude Code only by design. Dropping all three into one plugin marketplace does not work.
Self-reported metrics and what stays blank
Wherever scores appear, all three give only project self-reporting, with no third-party certification. Alibaba's README says that with the same underlying model it achieves significantly higher Precision and F1 than Claude Code while using about one ninth of the tokens, and openly notes that Recall is lower, a deliberate trade of recall for precision. Cloudflare says a single run found roughly half of what repeated runs found in total. Alibaba's benchmark details are reproduced too: 50 popular open-source repositories, 200 real pull requests, and 10 programming languages, cross-validated by more than 80 senior engineers with 1,505 annotated ground-truth issues, the dataset published on Hugging Face. These are self-reported metrics from each repository, reproduced here as such, never as independent measurement, and never used for cross-benchmark comparison or ranking. The blanks need stating too. Cloudflare's skill is only the single-repo starting point from which its vulnerability discovery harness was seeded; that harness later grew into a multi-stage, fleet-wide system, so calling the skill Cloudflare's only vulnerability discovery system is wrong. Alibaba and BugHunter both show a pushed date of 2026-11-16, but stars still follow the November 17, 2026 snapshot and are not mixed with live movement. What can be checked ends there.
FAQ
Q1: Can the three licenses be summarized in one line? A1: No. Cloudflare is a single MIT with copyright line 2025-2026 Cloudflare, Inc.; Alibaba is Apache-2.0; BugHunter is dual-licensed, MIT for code and CC BY 4.0 for content, with the GitHub API reporting only MIT. The three structures differ, and so do the commercial obligations.
Q2: Who performs independent verification? A2: Only Cloudflare. It uses fresh agents to verify final source claims and insists that the checker is never the finder. Alibaba's word independent refers to positioning and reflection modules, while BugHunter's triage-validation and 7-Question Gate are submission gates; neither is independent verification.
Q3: Is 83 skills a hard number for BugHunter? A3: Per the README it is 83 skills; the GitHub API description lags and says 82. This review follows the README and preserves the discrepancy. Of the 681 reports, only 433 are individually cited, so writing that all 681 are individually cited is wrong.
Q4: Do all three need an API key and work as Claude Code skills? A4: Neither. Cloudflare needs no API key and runs through the Skills CLI; Alibaba requires a configured model endpoint by default, with Delegation Mode removing the OCR key, and does not use a Claude Code subscription login; BugHunter needs no separate key for its core, with only the optional generator requiring an Anthropic API key. Their install entry points and namespaces are not interchangeable.
Q5: Does this review offer buying advice or a benchmark ranking? A5: No. The axis is the responsibility boundary behind who audits AI-written code, plus a comparison of three licenses. All scores are self-reported, so cross-benchmark comparison, buying advice, and performance ranking fall outside its scope.
Related reading
License comparison for open weights, Inference engine licenses and hardware, GitHub AI tools metadata comparison, and Lightweight agent model comparison make for adjacent reading, with angles different from this one.