- Javier Garza of Snyk argues that blanket AI bans push use onto personal devices and accounts, where security teams cannot see it.
- He urged explained rules, approved options and a daily-updated inventory of AI models, datasets and agents, since AI parts can change at any time.
Javier Garza of Snyk argues that banning AI tools moves their use out of sight instead of removing the risk. He made the case in a workshop on the AI Engineer show. In his view, staff who meet a ban, or a rule nobody explains, route around it. Security teams then cannot see what is in use.
What was said
Garza pictured a company that bars AI coding tools until its AI policies are written. He guaranteed that 80 or 90 percent of developers would still find a way to use them. That figure was his estimate, not a survey.
He also gave a field example. At banks and financial firms that limited AI tools, some staff kept a private laptop beside the work laptop. They ran the tool there and messaged the output to themselves.
He called this shadow AI, meaning unapproved use, "invisible by design." His verdict: "Just banning the AI use doesn't help." He added that "if you don't train people and educate, they will work around anything."
His answer to the blind spot is an AI bill of materials, an inventory of the models, datasets, agents and pipelines a company uses. A software bill of materials lists libraries and packages, which behave predictably. AI parts are non-deterministic, meaning they can change at any time, and models now arrive through model hubs, not package registries. So Garza said a one-time list fails and scans should run at least daily. Hidden tools can also leak in plain ways. He said some agent tools lack a secret manager, so users type an API key, a password for software, into the agent chat.
Why it matters
Our reading: a ban looks cheap because its cost arrives later. It lands on the security team, which must defend use it cannot see. An unlogged tool might hold a pasted API key, and nobody would know.
A ban without an approved alternative is not a control. Garza listed governance without alternatives among the approaches that fail. Offering sanctioned tools, explaining the limits and scanning often gives security something to defend.
The other side
Garza works with Snyk, which offers the inventory tool he demonstrated, so he has a stake in the remedy. His 80 to 90 percent figure was a guess, and the excerpts offer no measurement of how often bans fail.
He also showed no evidence that education works. He conceded that his own company's strict policy on an AI assistant is a pain for staff. The conversation did not settle what a bank should do when some data truly cannot go near an AI tool.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- AI Engineer: AI Security Engineer Foundations + Certificate — Javier Garza, Snyk (2026-10-11)
