Applications architect, @Amazon
Your AI agent has prod access. Your controls don't know it. Hot takes on AI agents, guardrails and what breaks in production.
I'll say it: "human in the loop" is mostly theater. Nobody reviews 40 agent approvals a day. They just click yes.
If autonomous agents are the new software architecture, the controls have to live where the action actually happens: the database, the API, the deploy.
@sagnik_ghosh_11 Exactly and CI gates and branch protection cover the code the agent writes. the harder part is what it does at runtime: the query it runs, the API it calls, that needs a check at the moment of execution too. What do you use for that layer?
the next big outage won't come from a hacker. It will come from your own agent doing exactly what you asked.
We are building systems where the caller decides what to do at runtime, and we still protect them with controls designed for humans clicking buttons.
Starlink ofrece internet estable y de alta velocidad para un uso doméstico con velocidades de hasta más de 230+ Mb/s. Streaming sin interrupciones, videollamadas, trabajo desde casa, juegos y mucho más.
Your AI agent has database access. What stops a DELETE without a WHERE?
Vericto is now in the official MCP Registry as com.vericto/sql-firewall. The agent calls check_sql before anything runs.
A real session: blocked, fixed, checked again.
we spent 15 years around teaching devs least privilege, code reviews, change approvals... and then gave an AI agent prod credentials and a prompt that says "please be careful" 😅
That's not architecture, bro. That's hope.
One line in a GitHub workflow:
- uses: vericto/check-sql@v1
It parses the .sql files a PR changes and checks them against your rules. DELETE without WHERE: the job fails, with a SARIF annotation on the line.
Auth with GitHub OIDC, so no stored secret.
An agent forgets the WHERE. Here's what happens when its connection goes through Vericto.
Real proxy, real Postgres, 1,000 rows. No WHERE: blocked. WHERE 1=1: blocked. WHERE id = 42: one row deleted.
999 left. Only the typing is scripted.
#AIagents
When an agent runs SQL on its own, nobody reads the query before the database does.
You can't allowlist that. The agent writes new queries on the fly, and a list of approved ones either blocks real work or gets loosened until it means nothing.
Check structure instead.
A coding subagent set out to test a reset routine. It wrote DELETE FROM for every table and ran it. Nobody approved it.
The .env had production credentials. A real customer's database was gone.
The repo had a guardrail. The script loaded the env on its own and skipped it.
The developer's public postmortem: https://t.co/PCosACXmGF
At the proxy, DELETE FROM users; gets BLOCKED · VERICTO-001 before the database runs it.
That only holds if production accepts connections from the proxy alone. Otherwise it's one more guardrail a script can skip.