Building infrastructure that actually works is hard enough; getting it noticed by judges who know their stuff is harder. Developer Aalishan001 just proved that a focused, open-source security tool can still cut through the noise. At the FORK. BUILD. MERGE. hackathon hosted by GDG Cloud Bengaluru on October 24, 2026, his project was selected as one of only ten winners out of 211 pull requests submitted that day.
The Problem With Shared Agent Memory
Most AI agents today share long-term memory in a way that is dangerously leaky. Imagine a hotel where every key opens every door, and privacy is just a polite sign on the wall. That is how vector tables often work: every user's data sits together, separated only by a WHERE clause the application developer must remember to write perfectly, every single time. If the similarity search ranks by distance rather than ownership, a query about your roadmap can accidentally pull up a colleague's private notes. No malicious attacker required.
Moving the Boundary to the Database
The winning solution, adk-memshield, shifts the security boundary from the application layer to PostgreSQL itself. It uses FORCE ROW LEVEL SECURITY with a database role that cannot bypass restrictions. This means a query with a missing or incorrect filter returns fewer rows, not someone else's rows. The companion tool, adk-leak-arena, is a red-team harness that runs two real ADK Runners side-by-side to attack both protected and unprotected setups. It proves that while unprotected pgvector tables leaked data in 97.5% of cross-boundary attack queries, adk-memshield leaked zero.
The Bug That Looked Like a Pass
Two days before the event, an adversarial review found a subtle failure mode. A malformed request with a NUL byte inside a spoofed user ID returned an HTTP 200 with zero rows. It looked like a security pass, but it was actually a crash. The database driver refused the parameter, the ADK tool caught the exception and logged a warning, and the arena counted it as zero leaks. The fix involved rejecting NUL bytes at the API level and updating the arena to report such cases as retrieval errors rather than successes.
Key Takeaways
- Defense in Depth is Real: Application-level filters are fragile; database-enforced Row Level Security (RLS) provides a robust safety net for multi-tenant AI agents.
- Benchmark Your Security: The
adk-leak-arenatool provides a reproducible way to test memory isolation, showing a 97.5% leak rate for unprotected setups versus 0% for RLS-protected ones. - Watch for Silent Failures: Error handling can mask security flaws. A crash that returns empty results can look like a successful block if telemetry isn't carefully monitored.
The Bottom Line
If you are building agents that share memory, stop trusting your app code to write the right WHERE clause. Push the boundary down to the database, test it with a red-team harness, and don't let silent errors look like security wins.