DCRH — Data Flow
Part of DCRH — Map of Content. Full Mermaid source files live in the
project folder: diagrams/01_architecture.mmd, diagrams/02_data_flow_pipeline.mmd,
diagrams/03_data_flow_dnr.mmd.
Level 0 (context): what crosses the trust boundary
flowchart LR
U["Operator / Claude Code session"] -->|target dir + optional THREAT_MODEL.md, config.yaml| ORC["vuln-pipeline orchestrator\n(trusted, unsandboxed)"]
ORC -->|build once, normal network| IMG[("ASAN Docker image")]
ORC -->|spawn, isolated| AG["claude -p agent containers\n(gVisor, egress: api.anthropic.com only)"]
AG -->|crash bytes only| ORC
ORC -->|results/, reports/, PATCHES/| U
AG -.->|API calls only| API[("Claude API")]
Level 1: the vuln-pipeline loop
flowchart TD
T["targets/<name>/ + Dockerfile"] --> B["Build ASAN image"]
B --> R{"--auto-focus?"}
R -->|yes| RC["Recon agent\nproposes focus_areas"]
R -->|no, config.yaml has focus_areas| F
RC --> F["Find agents (N parallel, isolated containers)\ncraft input until 3/3 crash"]
F -->|crashing input file| FB[("found_bugs.jsonl")]
F -->|PoC bytes only| G["Grade agent, fresh container\nre-verify crash is real"]
G -->|verdict| RES[("run_NNN/result.json")]
G -->|passed + --stream| J["Judge agent, serial\nnew / cleaner dup / skip vs manifest.jsonl"]
J -->|accepted or replaced| REP["Report agent\nexploitability write-up from PoC+source only"]
REP --> REPG["Report grader\nchecks claims are evidence-backed"]
REPG -->|scored| RJ[("reports/bug_NN/report.json")]
RJ --> P["Patch command\nfix agent drafts diff"]
P --> PG["Patch grade loop\nbuild + PoC no longer crashes + tests pass + re-attack"]
PG -->|verified fix| PATCH[("PATCHES/ diff")]
Level 1: the dnr-pipeline (detection & response) mirror
flowchart TD
LOGS["generate_logs.py --seed N\n(local-only, gitignored)"] --> HC["Hunt container\napp source + log corpus, read-only\n(NO answer key / grader / generator by image construction)"]
HC -->|INCIDENTS.json| GC["Grade container, fresh\nanswer key + scorer, read-only"]
GC -->|independently re-fires each PoC\nagainst a freshly seeded app| VERDICT[("result.json + scored write-up")]
VERDICT --> TR["/triage → /patch\n(same loop as static track)"]
Artifact glossary (see DCRH — Glossary for full definitions)
focus_areas— recon's attack-surface partition (YAML).found_bugs.jsonl— running ledger of submitted crashes, shared across parallel find agents so they can check for duplicates before adding to it.run_NNN/result.json— one find+grade run's verdict, written as soon as grading finishes.reports/manifest.jsonl— judge's dedup ledger across the whole batch.reports/bug_NN/report.json— final exploitability write-up + grader score for one unique bug.PATCHES// patch diff — verified candidate fix.INCIDENTS.json— D&R hunt/respond output: verdict, blast radius, proposed response per incident.
What never crosses a boundary
- The find agent's reasoning never reaches the grader — only the PoC bytes. This is what makes grading meaningful rather than rubber-stamping.
- The D&R hunt container never contains the answer key, grader, or log generator — enforced by what gets baked into its image, not by convention.
- Agent containers cannot reach the internet except the model API
(
api.anthropic.com:443by default, or the Bedrock/Vertex equivalent) — enforced by thevp-internalDocker network having no default route out, with a proxy container as the only egress path.
Related: DCRH — Architecture (HLD) · DCRH — Sandboxing and Trust Boundaries.