Week 43: the agent boundary, gateway checklists, and a supply-chain clock
This week: why agent security is an authorization problem, a gateway review you can run before the next deploy, and the first hour of a package incident.
On this page
The theme this week is boundaries. An agent is not secure because the model refuses; it is secure because the tool cannot do more than the request allows. A gateway is not secure because it has a policy page; it is secure because the caller cannot route around it. A supply-chain response is not finished when the package is removed; it is finished when the identity is rotated.
Lead
Agentic AI changes the threat model, not just the prompt
The model is the planner, not the security boundary. Once tools are involved, the assets that matter are tool identity, context, and side effects. The post includes a review checklist for a new agent and the uncomfortable question to ask before adding another tool.
Security
The first 60 minutes of a supply-chain incident
A practical order of operations: freeze deploys, snapshot evidence, revoke the highest-reach credentials, scope every install, then rebuild from a clean cache. The checklist is designed to be usable while the incident channel is noisy.
Securing LLM gateways: a production checklist
Identity, request policy, data handling, egress, observability, failure behavior, supply chain, and incident readiness. Updated this week with egress and token-scope checks.
AI
Open-weight models: the bill arrives after the benchmark
The API-versus-GPU comparison usually omits utilization, operations, evaluation, and the engineer who owns the upgrade. Use the shadow-run playbook before moving a customer-facing path.
Engineering
Platform engineering learning path: what to learn, in what order
A six-month, dependency-ordered path with an artifact at each phase. The last phase is operations, not another tool. Updated with “build one thing” milestones.
Watching
- MCP gateways are moving from demo to procurement. The auth model is still the weak point.
- Post-quantum TLS is deployable; certificate inventory is the long pole.
- Agent sandboxes are the right idea, but the escape-hatch story is not settled.
From the archive
Next.js middleware bypass: patch 15.2.3 now
Still the clearest example of why authorization cannot live only in middleware. If you missed it, check the affected versions and the header mitigation.
Corrections? Send them to [email protected].
Related reading
3 postsDiscussion
Spotted an error? Send corrections to [email protected] .