SECHIVE / OPERATOR CONTROL PLANE REV 04.A · 2026.05
§ 01 — Operator-driven, and defensible.

Findings that survive review,
not just findings that arrive fast.

SecHive is an operator-driven security research workbench under active development. You define the scope, hold the approval gates, and make the call on every finding while the runtime organizes hypotheses, validation notes, and reproducible evidence across pentest, bug bounty, source audit, mobile, and reverse-engineering modes.

Local-first
Proof-first
Operator-grade
Fig. 01 · Proof loop
I
Scope & Authorization
targets · exclusions · approval gates
II
Recon & Signals
routes · APIs · source · APKs · credentials
III
Skill Routing
specialist agents per signal class
IV
HypothesisGraph
scored · branched · refutable candidates
V
Runtime Validation
benign PoC · scope guard · negatives retained
VI
Proof Pack
artifact hashes · replay notes · redaction
VII
Reviewed Report
mode-specific · human-reviewed · ↺ loop
§ 02 — Public research record

Twenty-one public cases. Fourteen CVEs, four kernel fixes.

These are public advisory records and upstream commits, not benchmark scores. Twenty-one distinct findings across npm, PyPI, NuGet, Java, C/C++ and WordPress, plus four accepted Linux mainline contributions — each independently verifiable through the linked source. Five cases carry both a CVE and a repository advisory, so the identifier totals overlap and are never added together.

Verified as of 5 September 2026. Six public cases are listed at advisory level only, because their coordinators still withhold the technical detail. Further work remains in coordinated disclosure and is withheld in full — no identifier, product or technique — until the coordinating organization publishes.

21
Distinct public cases
A paired CVE and GHSA counts once.
14
Publicly disclosed CVEs
Published identifiers with public advisory records.
12
Published repository GHSAs
Maintainer-published advisories; five also carry a CVE.
4
Accepted Linux mainline fixes
Two authored patches, two reported-by credits.
vm2 · sandbox9.9 Critical

Read-only is not capability-safe.

An allowlisted node:test builtin came back wrapped in a read-only proxy. It still exposed a method that launches a Node process with caller-controlled interpreter arguments.

AdvisoryGHSA-qhwx-74w5-xhxq
WeaknessCWE-693
Fixed in3.11.7
Read the write-up
Apache Fory · Java9.8 Critical

Strict mode still trusted the wrong interface.

Strict class registration rejected an ordinary unregistered lambda capture class — and admitted the equivalent interface, because "no default methods" was read as proof the type was inert.

CVECVE-2026-64606
WeaknessCWE-502
Fixed in1.4.0
Read the write-up
vm2 · sandbox9.8 Critical

The guard counted wrappers, not capabilities.

A dangerous host prototype mutator was blocked directly and one layer deep. JavaScript lets the same capability survive arbitrary wrapper composition.

CVECVE-2026-47698
WeaknessCWE-913
Fixed in3.11.6
Read the write-up
python-statemachine · PyPI9.1 Critical

Reads were confined. Writes were not.

The restricted evaluator refused to read __class__. The assignment path walked the same segment as an intermediate destination and mutated shared class state.

AdvisoryGHSA-v3qq-3xvg-m77g
WeaknessCWE-915
Fixed in3.2.1
Read the write-up
isomorphic-git · npm8.8 High

The name was checked. NTFS disagreed.

Reserved-name filtering rejected .git, git~1 and trailing-dot variants — and accepted the one spelling NTFS resolves back to the real gitdir.

CVECVE-2026-77355
WeaknessCWE-22
Fixed in1.38.7
Read the write-up
Linux · net/sctpMainline

Generic TLV validity is not type validity.

A header-only SCTP parameter passed the generic walker. A typed consumer then read a 32-bit field that was never sent — and returned it to the peer in the state cookie.

Commit74b21f52c5c5
RolePatch author
MainlineLinux 7.2
Read the write-up
Disclosure boundary. Discovered and reported by Charles Vosburgh; where an advisory credits multiple reporters, the write-up preserves that shared attribution, and kernel pages separate reporting credit from patch authorship. Research was SecHive-assisted through source mapping, hypothesis generation, runtime construction, and skeptical validation; final review and disclosure ownership remained human. Reserved and embargoed work is withheld in full — no identifier, target, component, or technique — until the coordinating organization publishes.
Reference mappings
NIS 2EU 2022/2555 DORAEU 2022/2554 SOC 2AICPA TSC ISO 27001A.8.8 / A.8.29 PCI DSS 4.0Req. 11 HIPAA§164.308(a)(8) OWASPAPTS / AISVS
All frameworks →
Internal evaluation boundary. The figures below are project-reported workflow and purpose-built benchmark measurements. They are not CVEs, vendor-accepted findings, certification results, or independent validation.
94.23%
Internal XBOW black-box · 104 cases
Best-of black-box wins across the full recorded campaign.
100%
Internal source-aware run
Project-reported result on the same purpose-built case set.
104
Recorded benchmark cases
Internal workflow count, not confirmed vulnerabilities.
111
Juice Shop challenge tree
Purpose-built training target used for internal evaluation.
§ 03 — The market

Most AI security tools are good at producing activity. SecHive is built around a stricter question.

Scanners give you a long list of maybes. Agentic frameworks give you a long transcript. Both leave the human operator to reconstruct what happened, what was actually proven, and what is safe to put in a report.

SecHive collapses that gap. The operator stays in the chair — scope, approvals and disposition are yours — while the runtime writes its own audit trail as it works, so the report is the artifact, not the starting point.

Common workflow failure modes

Raw output without enough review context.

  • You get raw output — then spend hours rebuilding what actually happened.
  • No deterministic way to reproduce a finding on demand.
  • Evidence is scattered across screenshots, logs, and chat windows.
  • Negative results are dropped — you never see what the agent ruled out.
  • Provenance and reviewer decisions are often reconstructed after the run.
  • No policy gate on exploitation — easy to drift outside scope.
SecHive

A defensible report — built as the work happens.

  • Research tooling is designed to capture provenance while the work happens.
  • Validated findings can include deterministic replay notes and retained evidence.
  • Artifact hashes make later review and change detection easier.
  • Negative evidence is retained — refutations are first-class artifacts.
  • Structured review bundles keep evidence, disposition, and remediation together.
  • Exploit execution and external actions remain behind human approval gates.
§ 04 — The SecHive Loop

Seven stages between a target and a defensible report.

SecHive is not a single prompt, and it is not a fire-and-forget agent. It is an operator-controlled pipeline that enforces the scope you set, routes signals into specialist skills, plans hypotheses, validates them at runtime, preserves negative evidence, and only promotes a candidate to a finding once it clears your gate.

I — Scope
Authorization
Targets, surfaces, exclusions and approval gates loaded as policy.
II — Recon
Signals
Routes, APIs, source, APKs, browser state and credentials inventoried.
III — Routing
Skills
Specialist skills selected per signal. No prompt-of-everything.
IV — Hypothesize
Plan
Candidate weaknesses scored, branched and queued in HypothesisGraph.
V — Validate
Runtime PoC
Benign exploit attempted under scope guard. Negative results retained.
VI — Promote
Proof Pack
Evidence bundled with hashes, replay metadata, reviewer disposition.
VII — Review
Report
Mode-specific report rendered with artifact hashes, provenance, and reviewer disposition.
§ 03.A — Why a loop, not a prompt

The loop is the product.

Anything in security can be made to look impressive in a transcript. SecHive's loop is designed to preserve the work behind a finding: scope, evidence, validation steps, negative results, and a clear human review decision.

Walk through the platform
§ 05 — Operator UI

Mission control, running on your machine.

SecHive is local-first. The operator UI shows running campaigns, hypothesis graphs, evidence drawers, replay buttons and approval queues — all served from the same machine that holds your scope.

SecHive — Mission Control — localhost:7731
Campaigns
Active runs
All campaigns
New campaign
Evidence
Proof packs
Replay queue
Settings
Scope policy
Active campaigns
3
Running
14
Hypotheses
7
Validated
2
Promoted
api-auth-audit · prod-api.target.invalid
SK.02 bizlogic_hunter · SK.03 api_security · 6 hypotheses queued
pentest
00:23:41
running
bbp-submission-04 · hackerone-program
SK.01 recon_router · validation stage · 2 candidates pending
bug-bounty
01:04:17
running
pr-audit-#2847 · github/org/repo
SK.04 validation_bypass · diff-aware · 3 candidates
pr-audit
00:07:02
complete
mobile-apk-review · com.target.app v2.4.1
SK.07 apk_inspector · scheduled · awaiting authorization
mobile
queued
Hypothesis graph · api-auth-audit
High · SK.02
Signed authorization nonce not consumed — replay feasible via /v1/exec
Med · SK.03
IDOR on user_id param — numeric enumeration, no ownership check observed
Med · SK.04
Input validation denylist gap — unicode normalization bypass candidate on /submit
Low · SK.01
Staging endpoint exposed — no auth, returns 200 with schema leak
+2 more · negatives retained (9)
Read the platform
§ 06 — Run modes

One workbench. Six disciplines, each with its own truth.

Pentest, bug bounty, own-source bug hunt, mobile, reverse engineering and PR audit are different flows — different signals, different scope, different evidence standards. SecHive treats them that way instead of reusing one prompt with a different label.

  1. i. Pentest Authorized research with scope policy, approval checkpoints, and evidence organized for technical review. research mode
  2. ii. Bug Bounty Coordinated-disclosure reports with reproducible steps, CVSS context, and redaction-safe write-ups. research mode
  3. iii. Own-Source Bug Hunt Source-uploaded analysis where suspicions are kept distinct from validated runtime findings until proven. research mode
  4. iv. Mobile / APK Review Exported components, broadcast paths and binder surfaces — backed by static and runtime evidence. research mode
  5. v. Reverse Engineering Binary triage, anti-tamper inspection, and protocol reasoning where reproduction matters more than scan count. research mode
  6. vi. PR Audit Pull-request gating with diff-aware reasoning. Designed to catch regressions without flooding reviewers. research mode
§ 07 — Public proof

Public records and internal evaluations. Clearly labeled.

The public research record includes disclosed CVEs, published GitHub Security Advisories, and three accepted Linux mainline fixes. Purpose-built Juice Shop and XBOW-style runs are project-reported internal evaluations, not independent certifications.

SURFACE.APublic record

Published research

Independent, externally linkable records separate from work still moving through coordinated disclosure.

Public CVEs14
Published GHSAs12
Linux mainline fixes4 accepted
View the research record
SURFACE.BInternal evaluation

OWASP Juice Shop

Full unredacted black-box and white-box reports against the current 111-challenge tree. Routes, payloads, source references and evidence snippets retained.

Live runtime findings35 / 111
Source-aware items58 / 111
CVE references20
Read the report
SURFACE.CInternal evaluation

XBOW-style campaign

One hundred and four recorded validation cases, paired black-box and white-box. 98 black-box wins, 104 white-box wins, 100% any-win.

Recorded cases104
Black-box wins94.23%
White-box wins100%
Read the campaign
Claim boundary. Public advisory and upstream links are independently verifiable. Benchmark figures are project-reported results from controlled internal evaluations; they are not certifications, independent audits, public vulnerability totals, or vendor acceptance claims. Scoring methodology is documented per surface.
§ 08 — Compliance

Research artifacts mapped to familiar security frameworks.

The project documents how technical research artifacts can relate to SOC 2, ISO/IEC 27001, NIS 2, DORA, PCI DSS 4.0 and HIPAA. Framework references describe evidence mapping and control alignment. SecHive is not presently certified or independently audited under these frameworks unless explicitly stated.

§ 09 — Research pattern

A finding needs a reproducible broken invariant.

This generalized runtime-authorization replay pattern shows the evidence shape used to test a hypothesis without exposing confidential targets or embargoed work.

1

Hypothesis

A signed runtime authorization carries a nonce-like value. bizlogic_hunter flags the value as never observed in durable state.

2

First execution

The relayer submits the signed action. The validation function returns OK. The execution wrapper performs the privileged side effect.

3

Replay

The same authorization bytes are submitted again, unmodified. The nonce is not consumed; the digest validates.

4

Second execution

The privileged side effect occurs a second time. The invariant "one signed authorization authorizes one execution" is broken.

5

Promotion

After human review, the candidate can be promoted. The proof pack contains the signed payload, two execution receipts, side-effect deltas, the validation source path, artifact hashes, and a deterministic replay.sh.

Open full write-up
§ 10 — Output

A readable report. A practical manifest for engineering.

The workflow is designed to produce both a human-facing report and a machine-readable proof pack with provenance, a redaction manifest, artifact hashes, and replay material for technical review.

EXAMPLE — replay.sh in a proof pack
# sechive-proof-pack/replay.sh — sha256:7c3a…
set -euo pipefail
TARGET="https://lab.example.invalid"
AUTH="$(cat artifacts/signed-authorization.b64)"

# first execution — expected OK
curl -sS -X POST "$TARGET/v1/exec" \
  -H "X-Authorization: $AUTH" \
  -d @artifacts/payload.json | tee out/01.json

# replay — invariant broken if 200 OK
curl -sS -X POST "$TARGET/v1/exec" \
  -H "X-Authorization: $AUTH" \
  -d @artifacts/payload.json | tee out/02.json

diff <(jq .receipt out/01.json) <(jq .receipt out/02.json) \
  && echo "REPLAY ACCEPTED — finding confirmed" \
  || echo "replay rejected — invariant holds"
PROOF PACK CONTENTS
findings.provenance.jsonrecorded
review-disposition.jsonhuman
artifacts/ (raw evidence)sha256
replay.shdeterministic
redaction-manifest.yamlscoped
negative-evidence/retained
model-cost.jsonaccounted
report.md / report.htmlrendered
§ 11 — Ready

Bring an authorized scope. Leave with a defensible report.

SecHive is local-first. Pilots run against your own infrastructure or against a controlled lab — never against systems you don't own or aren't explicitly authorized to test.