What is SecHelix

A finding is a claim until it survives verification

SecHelix is an open-source Agent Skill and orchestration methodology for repositories and environments you are authorized to test. It coordinates code-reading agents, security tools, browser and runtime evidence, and an independent verifier under one shared standard.

Those three lines are why the workflow has a verification phase at all. A trusted finding has to establish attacker control, reachability, a failed security boundary, a bounded safe reproduction, concrete impact, a root cause, a fix, and regression proof. Anything short of that is reported as what it actually is.

The pipeline

1. ScopeRecord the target, execution mode, authorization, exclusions, and stop conditions before hunting.
2. MapBuild entrypoints, identities, assets, trust boundaries, and state machines from evidence.
3. SelectLabel every catalog hypothesis applicable, not applicable, unknown, or blocked.
4. HuntSpecialist lanes produce candidates. Scanner labels never become verdicts on their own.
5. VerifyAn independent verifier receives the claim and tries to refute it.
6. FixRepair the canonical invariant rather than each symptom.
7. RegressAdd a test that fails against the vulnerable control and passes after the fix.
8. RetestRe-run the original claim against the built application, not the source text.
9. Report and gateEmit the canonical report and a fail-closed release decision.

Four honest outcomes

Applicability, verification, and release each have their own closed vocabulary. Nothing in the workflow can quietly convert uncertainty into success.

Missing evidence is never treated as absence. An unconfirmed or partly unauthorized scope blocks execution; it does not make checks inapplicable.

What is different

Verification is first-class

Important candidates go to a verifier whose job is to disprove them. Compensating controls, unreachable states, missing attacker control, or impossible prerequisites are all valid reasons to reject a candidate.

Business logic is first-class

Security bugs live between individually valid actions: refund plus late provider success, delivery plus cancellation, cost edit plus finalized payout, two admins plus one assignment. Exact-once behaviour, state machines, and accounting truth are treated as security surfaces.

Runtime proof can outrank static confidence

A typecheck can be green while the browser flow is broken. Unit tests can be green while a database constraint or an authorization boundary behaves differently. Proof is required at the layer where the invariant lives.

AI-built code gets normal scrutiny

Built with AI is not itself a vulnerability class. The implementation is checked for missing server-side authorization, client-controlled identity and price fields, dynamic queries, unsafe HTML, SSRF, weak upload validation, permissive CORS, home-grown auth logic, missing replay controls, and overprivileged agent tools.

What it is not

• Not a scanner wrapper. Adapters normalize tool output into candidate evidence; they never assign a verdict.
• Not a hosted service. The framework is a portable skill plus contracts, and it runs inside your coding agent.
• Not a licence to test third-party systems. Authorization is an explicit prerequisite recorded in the scope.
• Not a benchmarked accuracy claim for the workflow. One blind label-only run is published; applicability, verification, regression-proof and release-gate performance remain NOT_MEASURED.

The framework is Apache-2.0. The methodology lives in SKILL.md, and every structural claim on this site is traceable to a file in that repository.

Where to go next