Containing the Agent, Counting the Value
By 2026, enterprise AI agents stopped advising and started acting. Eight articles on containment: when an agent does the wrong thing, can you stop it, can you prove it, and is it even worth running, ending in a value-and-risk ledger that tests every workload six ways and never blends them into one comforting score.
Three parts: A why it matters (1-2), B how to contain it (3-5), C the Value & Risk Ledger (6-8)
Platform & security architects, CISOs, AI risk & governance leaders
Read 1-8 in order: Part A for the case, B for the controls, C for the ledger
Reading path
Start at Part 1, or jump to the part you needWhy agent containment suddenly matters
In 2026, enterprise AI agents stopped advising and started acting, wired to tools, credentials and other agents.
What you're actually trying to contain
You're not defending against a VM escape; you're containing legitimate capability used through paths no one anticipated.
Software controls: stopping an agent in the request path
Which controls actually stop an agent, and which only tell you after the fact?
Hardware and out-of-band controls: stopping an agent from outside
Isolation is the baseline, not the answer.
Proving a control works: evidence, not assertion
A control is only as good as its most recent proof.
The six tests: why one score would lie to you
A single AI health score is exactly what lets a strong return hide a missing kill switch.
How to build the risk ledger
How do you run the six tests without rebuilding your observability stack?
What data to feed it, and where to start
Everything hangs on one join key, and on four inputs no platform will ever hand you.
Follow on LinkedIn to see each new article and series as it is published.
Other series
The Autonomous NOC
47 alarms to one root cause in under 90 seconds, safely enough to act on a live network. Twelve articles on how, why, and what it costs.
Bank Foundation Models
Why a bank should train a foundation model on its own event data, and how: from privacy-first data to production and Nova Forge.
GAD-P: Governed data access
Let every business user ask your data questions in plain English, with governance enforced by code rather than by prompt.
Dark Web Fraud Signals
Catch fraud techniques on dark-web forums weeks before they hit your transaction monitoring.
Real-time Fraud with TGNs
Graph-based fraud detection in under 100 ms that keeps working when you move to a new payment domain.
The Fast Path
Cut bank loan decisions from about 10 days to minutes, in six weeks.
AI Value Strategy
95% of AI projects never reach production. A seven-part guide to making sure yours delivers value you can measure.
Synthetic Data
How to trust synthetic data, and how to use it where real data is restricted.
Startup AI Readiness
Every AI readiness framework was built for enterprises. This 10-part, research-grounded one is for startups and teams under 50: assess where you stand, decide build vs buy, ship your first agent, measure it, govern it and price it.