Shipping AI Into 5 Major Banks Without a 12-Month Compliance Wait

Al Chen, Cisco17:42 · Oct 2026 · 384 views
Thumbnail for Shipping AI Into 5 Major Banks Without a 12-Month Compliance Wait Watch on YouTube
TL;DR
  1. 1

    Banks need evidence that vulnerabilities are fixed, data can be restored correctly, and sensitive information stays protected before they approve production AI.

  2. 2

    Agents can write exploits and tests inside isolated environments to check whether a CVE is still exploitable after a package or image is updated.

  3. 3

    Compliance work moves faster when customer problems are broken into engineering, workflow, and human and organizational work streams.

Summary

Al Chen explains how Cisco deploys AI observability software into production environments at five major banks. Standard SaaS deployment does not fit air-gapped banking systems because outbound internet access can expose sensitive information. Chen focuses on three areas: the build pipeline for moving images into bank clusters, disaster recovery and data correctness, and preventing PII or MNPI from leaking through internal systems such as LLM judges. He argues that version bumps and successful backup jobs are not enough. Teams need evidence, such as exploit tests, row-count checks, replica convergence checks, and controls over judge reasoning. Data residency also has to be designed into replication. Chen is candid that production timelines depend on infrastructure, security reviews, GPU procurement, image approval processes, and organizational alignment. His practical advice is to treat each customer problem as an engineering and operational problem, then show the work behind every compliance answer.

Key ideas
00:57

Production banking AI has to pass checks that prototypes avoid

Chen distinguishes production applications used by millions of bank customers from development prototypes, staging systems, and internal retrieval agents. Standard SaaS deployment fails in an air-gapped bank because it requires outbound internet access and could leak sensitive information. His team works with banks on the processes required to reduce production deployment time from a year to a couple of weeks. He organizes the work around the image build pipeline, regional data resiliency, and controls that prevent information from leaving the bank's VPC or reaching internal employees improperly.

02:32

CVE remediation depends on each bank's versions, scanners, and policies

A platform may contain many images, each with several CVEs, while different banks run different image versions and use different scanning tools. Banks also set their own severity rules and remediation SLAs. Some vulnerabilities come from transient dependencies in open-source projects that the engineering team cannot directly change, so the team has to explain those dependencies to bank security groups. Package upgrades and rebuilt base images can fix some issues, but a newer version may be unacceptable to one bank while an older version is required by another.

05:35

Exploit tests provide stronger evidence than a package version number

Chen describes a sandboxed harness with instructions, the application's Docker images, and tests that measure the agent's result. The agent receives a CVE and a supposedly fixed package version, writes an exploit based on the older version, and then writes a test to check whether the exploit still works. Multiple containers or sandboxes can run trials in parallel. This gives the engineering and security teams evidence that the exploit is no longer present in the application code, rather than only proof that a package was bumped.

07:03

Backup success does not prove that restored data is correct

Chen uses ClickHouse as an example of a database without managed cross-region failover and replication. A possible pattern has ClickHouse in a primary region, a job backing up to a bucket, and another job restoring the data in a disaster-recovery region. The backup tool can report that the jobs ran, but that status does not prove the data is current. Bad disks, bad nodes, migration problems, or replication lag can leave the restored data stale. Teams should compare row counts and check whether replicas within a region have converged.

09:35

Disaster recovery choices also decide where customer data resides

Replicating data to a bucket creates a data-residency decision because the records may cross jurisdictions. Chen describes one approach in which IDs and trace metadata can move between regions while the core trace data remains in its original region. The design has to account for jurisdiction laws, not only whether the backup and restore workflow completes. Data location is part of the architecture even when the team did not set out to make a residency policy.

10:27

An LLM judge can expose PII after the user interface redacts it

Redaction can happen at the SDK, gateway, or middleware layer, and the visible interface may appear clean. Chen warns that the system prompt for an LLM judge may ask for an explanation so a human can review the score. The judge's reasoning can then repeat or unredact PII or MNPI in a stored trace, even though the UI's customer information looks protected. Teams should inspect judge prompts and traces, not stop at the rendered interface.

11:36

Protecting judge reasoning adds checks and another model call

One design lets the judge read raw traces while preventing its reasoning from being persisted by default. The score can remain available for evaluations. Application code can decide whether a record contains sensitive information, showing reasoning only when it is safe. If banks need reasoning for sensitive records, a second judge can receive redacted input and produce reasoning that is safer to display. Chen states that zero leakage and data sovereignty have a cost because this approach requires another LLM call on the same request.

13:10

Compliance speed depends on operational preparation and organizational alignment

Deployment estimates stretch when teams lack image-pull processes, ready GPUs, formal security reviews, a custom model gateway, role-based access configuration, or an approval process for moving images into a registry. Chen says many delays are less technical than organizational. His approach is to show the evidence and work behind each answer, then divide customer problems into engineering, workflow, and human and organizational work streams. The production checklist extends beyond the three pillars discussed in the talk.

"I'm not talking about prototypes in dev or staging or even like internal RAG agents to increase internal employee productivity. I'm talking about the applications that these banks' millions of customers use every single day."00:57
Who should watch
  • You are shipping AI into regulated production systems and need a practical way to prepare for bank security and compliance reviews.
  • Your vulnerability process stops at package upgrades, or your disaster-recovery checks only confirm that a job completed.
  • You use LLM judges or trace-based evaluation and need to understand how reasoning can expose data that the UI appears to have redacted.