Plugin name
sechelix
Short description
Evidence-first application-security Agent Skill for repositories and environments you are authorized to test. Every candidate finding goes to an independent verifier whose job is to disprove it before it is reported.
GitHub repository
omarmohelal/SecHelix
Plugin path inside the repository
No response
Ref to review
v3.4.0-alpha.1
Commit SHA to review
f76d32cf54d30ff9862e674b817d92df77289406
Version
3.4.0-alpha.1
License identifier
Apache-2.0
Author name
Omar
Author URL
https://github.com/omarmohelal
Homepage URL
https://sechelix.com
Keywords
appsec
application-security
security-audit
security-review
agent-skills
copilot-cli
ai-security
mcp-security
authorization
business-logic
devsecops
Additional notes for reviewers
What it is. SecHelix maps the attack surface, selects applicable hypotheses from a 546-item catalog (21 families × 26 verification lenses), hunts across 17 specialist role profiles, and then sends every candidate to an independent verifier whose job is to refute it. Applicability resolves to APPLICABLE / NOT_APPLICABLE / UNKNOWN / BLOCKED, so missing evidence is never treated as absence. High and Critical findings require regression proof. The release gate is fail-closed: PASS, PASS_WITH_KNOWN_RISK, BLOCKED, or INCOMPLETE.
Coverage emphasis is business logic, payments, race conditions and idempotency, authorization (BOLA/IDOR), supply chain, and AI/agent/MCP security. There is also an UNTRUSTED_REPO mode in which repository content is treated as data and never as control instructions — CLAUDE.md, AGENTS.md, settings files and hooks inside a reviewed repository cannot grant a capability or widen scope.
Structure. 15 JSON Schema Draft 2020-12 contracts, 18 Gold Check Packs (including framework packs for Next.js, Express, Django, Supabase, Spring Boot and Laravel), 11 read-only scanner-evidence adapters, and 38 paired eval fixtures / 76 cases. Python standard library only.
Honest status, stated up front:
- The full SecHelix workflow benchmark is
NOT_MEASURED. Applicability accuracy, verifier effectiveness, regression-proof rate and release-gate accuracy have never been measured end to end.
- One blind label-only evaluation now exists, recorded 2026-09-02 in
evals/results/claude-sonnet-5-blind-2026-09-02.json. It is deliberately not presented as SecHelix performance: the 76-case synthetic run measured one label per file and did not execute the verifier, adapters, remediation, regression or release-gate workflow. Current evaluation report: https://github.com/omarmohelal/SecHelix/blob/main/docs/research/evaluation-report.md. No benchmark-superiority claim is being added, and no comparison to any other tool is offered.
- The repository contains a keyword baseline flagged
is_sechelix_result: false. It is a naive regex matcher used to validate the scoring harness and evidence fixture difficulty. It is not a SecHelix score.
- There is one published case study: an authorized owner self-audit of a small app, in which one MEDIUM finding was verified, fixed and regression-proved, and one plausible high-severity XSS candidate was independently refuted. It is a worked example of the process, not a performance claim.
- This is alpha software; contracts and interfaces can still change. Happy to be listed with a caveat, or to wait, if you would prefer the marketplace carry only stable projects.
On Copilot specifically: the repository ships .github/skills/sechelix/SKILL.md because GitHub documents that path as a repository skill directory. The project's own compatibility matrix records Copilot loading as DOCUMENTED, not VERIFIED — the path is right, but no Copilot session was observed loading it here. I did not want to overstate that.
I am the author of the project.
Submission checklist
Plugin name
sechelix
Short description
Evidence-first application-security Agent Skill for repositories and environments you are authorized to test. Every candidate finding goes to an independent verifier whose job is to disprove it before it is reported.
GitHub repository
omarmohelal/SecHelix
Plugin path inside the repository
No response
Ref to review
v3.4.0-alpha.1
Commit SHA to review
f76d32cf54d30ff9862e674b817d92df77289406
Version
3.4.0-alpha.1
License identifier
Apache-2.0
Author name
Omar
Author URL
https://github.com/omarmohelal
Homepage URL
https://sechelix.com
Keywords
appsec
application-security
security-audit
security-review
agent-skills
copilot-cli
ai-security
mcp-security
authorization
business-logic
devsecops
Additional notes for reviewers
What it is. SecHelix maps the attack surface, selects applicable hypotheses from a 546-item catalog (21 families × 26 verification lenses), hunts across 17 specialist role profiles, and then sends every candidate to an independent verifier whose job is to refute it. Applicability resolves to
APPLICABLE/NOT_APPLICABLE/UNKNOWN/BLOCKED, so missing evidence is never treated as absence. High and Critical findings require regression proof. The release gate is fail-closed:PASS,PASS_WITH_KNOWN_RISK,BLOCKED, orINCOMPLETE.Coverage emphasis is business logic, payments, race conditions and idempotency, authorization (BOLA/IDOR), supply chain, and AI/agent/MCP security. There is also an
UNTRUSTED_REPOmode in which repository content is treated as data and never as control instructions —CLAUDE.md,AGENTS.md, settings files and hooks inside a reviewed repository cannot grant a capability or widen scope.Structure. 15 JSON Schema Draft 2020-12 contracts, 18 Gold Check Packs (including framework packs for Next.js, Express, Django, Supabase, Spring Boot and Laravel), 11 read-only scanner-evidence adapters, and 38 paired eval fixtures / 76 cases. Python standard library only.
Honest status, stated up front:
NOT_MEASURED. Applicability accuracy, verifier effectiveness, regression-proof rate and release-gate accuracy have never been measured end to end.evals/results/claude-sonnet-5-blind-2026-09-02.json. It is deliberately not presented as SecHelix performance: the 76-case synthetic run measured one label per file and did not execute the verifier, adapters, remediation, regression or release-gate workflow. Current evaluation report: https://github.com/omarmohelal/SecHelix/blob/main/docs/research/evaluation-report.md. No benchmark-superiority claim is being added, and no comparison to any other tool is offered.is_sechelix_result: false. It is a naive regex matcher used to validate the scoring harness and evidence fixture difficulty. It is not a SecHelix score.On Copilot specifically: the repository ships
.github/skills/sechelix/SKILL.mdbecause GitHub documents that path as a repository skill directory. The project's own compatibility matrix records Copilot loading asDOCUMENTED, notVERIFIED— the path is right, but no Copilot session was observed loading it here. I did not want to overstate that.I am the author of the project.
Submission checklist