adk-leak-arena
v0.1.0
Red-team memory isolation inspector · Unprotected pgvector vs adk-memshield (PostgreSQL RLS)
DB: probing...
Ollama: probing...
Gemini: probing...
Engine
deterministic_agent (offline)
ollama_gemma
gemini_api
Seeded dataset
Run Full Benchmark
Reset / re-seed DB
Caller
top_k
1
3
5
10
20
seed_size
36 rows (baseline)
250 rows
500 rows
1000 rows
2000 rows
Strip all SQL WHERE filters (chaos)
Hardware Load Notice:
High-load stress parameters or local LLM inference selected.
Use [KILL RUN] inside the benchmark modal to abort and restore the 36-row baseline immediately.
Loading scenarios...
Run Side-by-Side Query
Ready to Run
RLS Write-Policy & CTF Lure Injector (/api/inject)
Test owner spoofing, clearance write-up escalation, and private/shared lure writes against PostgreSQL WITH CHECK
Acting Persona (Session Writer)
Claimed Owner ID (optional spoof)
Target Scope
private
tenant_shared
min_clearance (1..5)
1 (Analyst / Contractor)
2 (Lead)
3 (Executive / CFO)
4 (Board)
5 (Root Vault)
Canary Token
Fill Sample Rank-1 Lure
Execute Write Test
Performance & RLS Kernel Overhead
Unprotected SQL:
-
MemShield RLS SQL:
-
RLS Kernel Delta:
-
End-to-End Agent Turn:
-
Stage a scenario, then click Run Side-by-Side Query.
5-Scenario + S6 Boundary + Control Red-Team Evaluation Matrix
Copy Markdown Table
Download JSON
[x] Close
Live Benchmark Runner
Run History & Comparison
top_k
1
3
5
10
20
seed_size
36
250
500
1000
2000
repeats
1
3
5
label
Start Benchmark Job
KILL RUN
Ready to run evaluation suite.
0 / 0 (0.0%)
Raw Markdown
Run A
Compare with Run B
Compare Side-by-Side
Refresh History
Seeded multi-tenant memory dataset
[x] Close