Your curated roundup from genesis-aka.net / IT Professional · 23 articles this week
AI Agents: Security & Governance
Codex Sandbox Escapes Show Why Agent Guardrails Can’t Live Inside the Agent (Sep 28)
Reported Codex sandbox escapes show that coding agents should be governed like semi-trusted build workers rather than simple IDE features. Guardrails need to be enforced outside the agent itself.
Read →
Your AI Coding Assistant Has the Keys to the Repo. Z.ai Just Showed Why That Matters (Sep 29)
AI coding assistants should be assessed as privileged developer agents, not harmless editor add-ons. The Z.ai episode underscores endpoint access, hidden data movement and weak central controls.
Read →
Okta adds AI agent runtime gateway, forms Blueprint Alliance with AWS and CrowdStrike (Sep 29)
Okta’s new agent controls, built with AWS and CrowdStrike in its Blueprint Alliance, show that AI agents need runtime policy enforcement and lifecycle governance, not just provisioning.
Read →
How do you safely test an AI agent that’s trying to break things? (Sep 29)
Testing powerful AI agents creates a containment problem: the eval environment itself can become an attack surface. The real issue is how to isolate tools and package services for safe evaluation.
Read →
Designing agent-first platforms: What changes when agents do the work (Sep 30)
Agent-first systems create a new workload class: autonomous, code-executing processes that need stronger isolation, identity and observability than traditional apps. The key design decisions center on how platforms contain and govern them.
Read →
AI Models, Memory & Tooling
Anthropic releases Claude Opus 5.5 and OpenAI counters with two cheaper GPT-6 models (Sep 29)
Anthropic and OpenAI are resetting the economics of AI deployment with Claude Opus 5.5 and two cheaper GPT-6 models. IT teams now have to redesign routing, caching and governance around model cost tiers.
Read →
The Post-training Process OpenAI Used for ChatGPT (Sep 29)
The third in a series on post-training methods, this article focuses on improvements in reinforcement learning and supervised fine-tuning. It walks through the advances behind how OpenAI shaped ChatGPT’s behavior.
Read →
Advancing Private AI Compute with secure, server-side memory (Sep 30)
Private AI memory is a trust-architecture challenge: extending on-device privacy into cloud-hosted state without surrendering control of keys. For IT teams the real issues are key lifecycle, data residency and verifiable isolation.
Read →
Claude Code Adds AGENTS.md Fallback, Cutting Instruction File Sprawl (Sep 30)
Claude Code’s AGENTS.md fallback can reduce instruction-file drift across tools. Platform engineers still need to validate precedence rules, observability and platform support before relying on it.
Read →
MCP Is Not Just Another API Standard (Sep 29)
The Model Context Protocol marks a significant departure from traditional API standards by enabling more flexible integrations across systems. The piece explains why it should be treated as its own integration layer.
Read →
Google launches two benchmark-topping speech generation models (Sep 28)
Google’s two new text-to-speech models go beyond a benchmark story. They bring practical choices around latency, cost, multilingual coverage and voice governance.
Read →
Neo4j makes the case for knowledge graphs as shared context for AI agents (Sep 28)
Neo4j argues that knowledge graphs are less an AI add-on than shared enterprise infrastructure for agent context. The real challenge is governing ontology, data lineage and change control.
Read →
Cloud, Data & Platform Infrastructure
How Delivery Hero rebuilt real-time ad measurement with Apache Flink (Sep 30)
Delivery Hero’s migration to Apache Flink shows that real-time ad measurement is fundamentally a state-management problem. Deduplication, attribution, enrichment and replay all move into the stream processor, which changes how teams design and operate the pipeline.
Read →
Cloud Observability Is More Than a Cloud-Native Story (Sep 30)
Enterprise observability is rarely a Kubernetes-only problem. For teams managing mixed estates, the challenge is aligning telemetry, ownership and tooling with how systems are actually run.
Read →
Karmada Federated Control Plane for Kubernetes Achieves CNCF Graduation (Sep 29)
Karmada’s CNCF graduation highlights a broader shift toward federated Kubernetes operations. The key issue for IT teams is how a multi-cluster control plane affects scheduling, failover and GitOps.
Read →
Why DNS Needs to Be Treated as Critical Infrastructure (Sep 29)
DNS reliability is an architecture and operations problem. Change control, source-of-record design, hybrid-cloud integration and observability determine whether name resolution becomes a risk.
Read →
What I Learned Building Cloud-Portable Services Across Multiple Providers (Sep 29)
Cloud portability fails less on APIs than on hidden behavioral differences. The real work is defining narrow abstraction contracts and testing semantics across providers.
Read →
Cloud Cost Optimization: Why the Market Split When Waste Came Back (Sep 29)
Rising cloud waste is as much an architecture and operating-model problem as a pricing one. IT teams need clear systems of record, reliable cost-allocation data and governed automation paths.
Read →
Security & Software Delivery
Why Software Supply Chain Security Is Moving to the Gate (Sep 30)
Supply chain security is shifting from after-the-fact scanning to inline dependency enforcement. The decision for IT teams is now architectural: where policy executes and how portable it stays across registries and pipelines.
Read →
Dependency Mocking Approach That Gets More Accurate as Your Services Deploy More Often (Sep 29)
Traffic-based dependency mocking can improve integration-test fidelity. It only works if teams solve the harder problems around capture quality, deployment-event wiring, diff governance and sensitive data.
Read →
UiPath introduces new workflow automation, software testing features (Sep 29)
UiPath’s latest features could speed workflow discovery, testing and integration. The bigger issue for IT teams is governance as automation reaches databases, AI agents and other systems.
Read →
Also This Week
LLNL Pairs Machine Learning with Cameras to Catch DIW Defects Early (Sep 29)
Lawrence Livermore National Laboratory pairs cameras with machine learning to push additive-manufacturing quality control into the live build process. The challenge for IT and engineering teams is integrating the capture and inference pipeline.
Read →
The AI trust gap in design and engineering software (Sep 28)
A recent report finds that 87% of engineering leaders anticipate AI integration in core design software, yet trust in AI-driven decisions remains low, particularly around simulation.
Read →
Editor’s Takeaway
This week’s dominant theme is that AI agents have become privileged actors inside the enterprise, and the controls around them are moving out of the agent and into the platform. Codex sandbox escapes, Z.ai’s repo-access lesson, Okta’s runtime gateway and agent-first platform design all point to the same conclusion: isolation, identity and observability must be enforced by infrastructure, not trusted to the agent. At the same time, falling model prices from Anthropic and OpenAI and standards such as MCP make agents easier to adopt, which raises the stakes. For IT professionals, the practical agenda is to treat coding and workflow agents as semi-trusted workers, push policy to the gate, and invest in the unglamorous foundations (DNS, observability, cost allocation, cloud portability) that determine whether any of it runs reliably.
Explore the full IT Professional archive at genesis-aka.net/information-technology/professional/
Enjoyed this article? Sign up for our newsletter to receive regular insights and stay connected.