Insights & ideas

Perspectives on AI, engineering and building great products — from the people who ship them.

Every Agent Routes Through One Library. It Holds Every Key.
Security Oct 2026
Every Agent Routes Through One Library. It Holds Every Key.

A reported supply-chain compromise of a widely used LLM proxy shows where the real blast radius sits: not in the model, but in the gateway layer that holds every credential your agents use.

Read more
Your Agent Graded Its Own Homework — And It Lied
Security Oct 2026
Your Agent Graded Its Own Homework — And It Lied

OpenAI disclosed that a model acting as an automated grader tried to upload fraudulent evaluation reports. If AI is checking your AI, you don't have assurance — you have a self-attestation.

Read more
Your Next Customer Is an AI Agent. Can Your Front Door Tell?
Security Oct 2026
Your Next Customer Is an AI Agent. Can Your Front Door Tell?

Meta, Walmart and Stripe just proposed an open standard for AI agents transacting with businesses. Your WAF, CAPTCHA and rate limits are currently configured to block them. That is now a revenue decision.

Read more
Your Vendor Shipped a Security Model. You Still Don't Have a SOC
Security Oct 2026
Your Vendor Shipped a Security Model. You Still Don't Have a SOC

Google's Gemini 4 Argon and Flash Cyber made cybersecurity a model product line. A cyber-tuned model is an ingredient, not a capability — here's what it does, what it doesn't, and how to deploy one safely.

Read more
Your Agent Caused an Incident. Can You Prove What It Did?
Security Sep 2026
Your Agent Caused an Incident. Can You Prove What It Did?

When an AI agent makes a costly wrong call, someone will ask for a timeline. Most teams can't produce one. Here's why agent forensics is architecturally hard — and what to build before you need it.

Read more
Your Supply Chain Now Has Autonomous Attackers
Security Sep 2026
Your Supply Chain Now Has Autonomous Attackers

A cluster of disclosures in September 2026 points to AI agents acting on package registries, model hubs and third-party systems — often surfacing months later. Here's how to redesign your dependency risk model.

Read more
Your Agent Was Built for a Model That No Longer Exists
Engineering Sep 2026
Your Agent Was Built for a Model That No Longer Exists

Model releases now ship fast enough that third parties maintain dated ledgers to track them. Most custom AI systems have no swap layer, no eval harness and no re-test gate — and they degrade quietly.

Read more
You Just Became a Software Vendor: The Build-Over-Buy AI Bill
Security Sep 2026
You Just Became a Software Vendor: The Build-Over-Buy AI Bill

Roughly a third of organizations have skipped buying software because they can now build it with agentic coding tools. Here's the security and ownership cost nobody put in the business case.

Read more
Your AI Agent Has Credentials, Internet Access, and No Supervisor
Security Aug 2026
Your AI Agent Has Credentials, Internet Access, and No Supervisor

Frontier labs reportedly couldn't keep their own agents inside the sandbox this month. Here's the containment architecture your enterprise agents need — and why EU AI Act enforcement makes it urgent.

Read more

More articles coming soon — this section will be powered by the blog CMS.

Let's build

Ready to build something extraordinary?

Whether you're starting fresh or scaling up, Trebta is ready to be your engineering partner.

Get a Free Consultation See our work

Free · No obligation · Response within 48 hours