decx-framework-vulnhunt

DECX Framework Vulnerability Hunting

Goal: discover exploitable vulnerability chains in the target framework to enable timely remediation and reduce business security risk. Identify proven, exploitable paths from Binder entrypoint to system-visible impact, validated through the evidence gate. Every intent created should serve this goal.

Evidence graph: observations as facts, relations as edges, investigation steps as intents.

Route elsewhere: decx-app-vulnhunt (APK app-layer), decx-cli (command help), decx-report (reports), decx-poc (PoC).

Main Agent: Delegate and Decide

Main agent does only these, nothing else:

1. init project + decx ard framework open final processed JAR
2. Delegate intent goals to subagent (one goal per delegation)
3. Read subagent summary + run chains to verify evidence paths
4. Decide promotion against references/risk-rating.md
5. Handoff to report/poc

Main agent must NOT run decx code, decx ard, read source code, load references, or analyze vulnerability patterns. All analysis happens in subagent.

Subagent: Owns Full Intent Lifecycle

Subagent receives: intent goal text + DB dir + DECX port. It owns the full lifecycle:

1. Create the intent via intent CLI (knows phase, references facts it discovers)
2. Run decx code/decx ard queries (must view code via DECX commands — never guess)
3. Load references/index.md for routing, references/patterns/*.md for trace cues
4. Write fact/edge directly to DB
5. Call solve --conclude <factId> or --fail when done
6. Return a structured JSON summary (contract below)

Subagent creates the intent itself (not main agent), so it always knows the correct --from fact IDs — no mismatch errors.

Subagent Summary Contract

After solve, subagent returns this JSON as its final output:

{
  "intent_id": "i003",
  "status": "done",
  "phase": "trace",
  "goal": "Trace Binder call to privileged sink",
  "facts": {
    "total": 6,
    "types": { "service-entrypoint": 1, "binder-reachability": 1, "control": 1, "identity": 1, "sink": 1, "impact": 1 },
    "key_fact_ids": ["f004", "f009"]
  },
  "has_proving_path": true,
  "result_fact": "f009"
}

DB is always source of truth. Summary is a routing aid — main agent uses it to decide next intents. Before promotion, main agent still verifies graph structure with chains.

Do not specify subagent type when delegating — just give the task goal + DB dir + DECX port.

Commands

# Main agent
node scripts/decx-analysis-db.mjs init   <dir> --session <name> --kind android_framework
node scripts/decx-analysis-db.mjs intents <dir> [--status <open|done|failed>]
node scripts/decx-analysis-db.mjs facts   <dir> [--prefix <type>]
node scripts/decx-analysis-db.mjs chains  <dir> [--root-prefix <type>] [--leaf-prefix <type>]
node scripts/decx-analysis-db.mjs export  <dir>

# Subagent
node scripts/decx-analysis-db.mjs intent  <dir> [--from <factId,...>] --goal "<question>" [--priority <n>] [--phase <stage>]
node scripts/decx-analysis-db.mjs fact    <dir> --prefix <type> --body "<text>" [--evidence <path>] [--confidence <0-1>]
node scripts/decx-analysis-db.mjs edge    <dir> --from <factId> --to <factId> --kind <proves|enable|carry|amplify|bypass|observe>
node scripts/decx-analysis-db.mjs solve   <dir> <intentId> --conclude <factId> | --fail "<reason>"
node scripts/decx-analysis-db.mjs chains  <dir> [--root-prefix <type>] [--leaf-prefix <type>]
node scripts/decx-analysis-db.mjs path    <dir> --from <id> --to <id>
node scripts/decx-analysis-db.mjs ancestors  <dir> --fact <id>
node scripts/decx-analysis-db.mjs descendants <dir> --fact <id>

Dispatch Protocol

Main agent follows this exact loop, nothing more:

1. init + decx ard framework open
2. Delegate to subagent: goal="Collect framework surface: Binder services, AIDL methods, system services", phase=surface
   Subagent: intent → analyze → write facts → solve → return summary
3. For each service-entrypoint fact in summary: delegate to subagent goal="Trace <factId> to privileged sink, prove 6-tuple", phase=trace
   Subagent: intent --from <factId> → analyze → write facts → solve → return summary
4. For each promoted pair: delegate to subagent goal="Validate composition edge <f1,f2>", phase=compose
   Subagent: intent --from <f1,f2> → analyze → write facts → solve → return summary
5. Check chains, promote against risk-rating.md, handoff

One intent = one subagent invocation. Do not batch multiple intents in one delegation.

Evidence Gate (Framework 6-tuple)

Promote only when ALL six fact types exist, connected by proves edges:

Type	What it proves
service-entrypoint	Binder service exposes method
binder-reachability	unprivileged app can reach the method
control	attacker-controlled parameter reaches sink
identity + permission-guard/appop-guard/user-guard	caller identity and authorization at trust boundary
sink	privileged operation reached
impact	system-visible consequence

Composition Edges

Edge	Meaning
enable	A creates access needed by B
carry	A transports UID/package/user/URI/Intent/token/Binder state to B
amplify	A increases B's framework/system impact
bypass	A defeats permission/app-op/identity/user/callee defense
observe	callback/broadcast/result/launched Intent makes B's impact visible

Rules

Rule	Why
Main agent never runs decx code/decx ard or reads source	main agent context stays clean
Subagent owns full intent lifecycle (create → analyze → solve → summarize)	subagent knows fact IDs, avoids --from mismatch
Subagent views code via decx code/decx ard commands only — never guesses	guessing produces hallucinated evidence
Subagent returns structured JSON summary after solve	main agent gets concise view without extra DB reads
DB is source of truth; verify with chains before promotion	summary is routing aid, not proof of evidence gate
One intent per subagent invocation	prevents context explosion
evidence must be file path	prevents token blowup
Method/service names are not evidence	registration ≠ reachability
Identity/authorization evidence required at trust boundary	framework findings need caller identity proof
Composition must be validated by dedicated subagent	adjacency ≠ composition
Cannot state poc-validated / runtime-validated	static analysis limit
Handoff at 60% context: export + session name	prevents truncation
Full method signatures: "pkg.Class.method(param):return"	precise identification
One final framework target per hunt; use collect/process/run path for split jars	intermediate jars lose AIDL/Stub/identity paths

References

• references/index.md — routing matrix (subagent loads this)
• references/risk-rating.md — severity levels (main agent loads before promotion)
• references/patterns/*.md — vulnerability shapes (subagent loads on signal)