What this video shows
Chen describes deployment work across five banks and focuses on checks that product demos omit. Teams must triage vulnerable images, prove that restored data is correct, keep sensitive information inside approved boundaries, and show reviewers how evidence supports each claim. One example shows how an LLM judge can repeat personal or market-sensitive information even when the visible application redacts it.
Convert every assurance into an executable check. Inventory images and dependencies, prioritize known exploitation and actual exposure, restore backups into an isolated environment, compare records, and send canary secrets through every trace and judge path. Store the evidence reviewers need without copying sensitive production data into a second uncontrolled system.
Use CISA's Known Exploited Vulnerabilities catalog as one risk-prioritization input rather than treating every CVE equally. NIST's Privacy Framework provides a broader process for identifying and managing privacy risk.
What you will learn
- A passing backup job does not prove restored records are complete and usable.
- Redaction must cover prompts, traces, graders, logs, and reviewer interfaces.
- Vulnerability severity needs asset, exposure, exploit, and remediation context.
- Compliance reviews move faster when engineering produces reproducible evidence.
How to apply this safely
- Map every model, judge, trace store, image, data path, and regional boundary.
- Automate image inventory and prioritize exploitable vulnerabilities using approved risk criteria.
- Run restore drills, compare checksums or business records, and test application reads from the restored copy.
- Seed canary identifiers and verify they never appear in disallowed outputs, logs, judges, or support tools.
Important limitations
- The five-bank experience and deployment timing come from the speaker and are not an independent benchmark.
- Requirements differ by institution, jurisdiction, data class, and deployment model.
- Automated exploit generation can create offensive artifacts and needs strict authorization and isolation.
- A second model used for redaction adds cost and another component that can fail.
Sources to check
- CISA Known Exploited Vulnerabilities Catalog CISA's authoritative list of vulnerabilities known to be exploited in the wild.
- NIST Privacy Framework A voluntary framework for identifying and managing privacy risk.
- Shipping AI Into 5 Major Banks AI Engineer video published October 20, 2026.
Continue learning on Learnetto
AI evals guide
Turn compliance claims into repeatable checks.
Sandbox AI coding agents
Constrain tools, credentials, and network access.
Verify AI citations
Check compliance claims against their evidence.