cmdev Academy — Honest AI & Cloud Security Writing

101 articles

SLSA Levels in Practice: What Most Organisations Claim vs What Their Pipeline Actually Enforces

Most organisations claim SLSA Level 2 or 3 on their security questionnaires. Most pipelines enforce Level 1 with manual sign-off. This piece walks the four SLSA levels honestly, names the gap between claim and enforcement, shows the reference toolchain for a defensible Level 2, and explains why the missing deploy-time verification step is where most claims fall apart.

June 16, 2026cmdev12m

Cloud Security

Cloud Security Posture Management in 2026: The Honest Vendor Comparison Wiz, Lacework, Orca, Palo Alto, CrowdStrike, and Aqua Won't Write

Every CSPM vendor comparison piece you have read is either marketing or written by a competitor with an axe to grind. cmdev has no commercial relationship with any of these vendors. This is the structurally honest version — six vendors, five dimensions, and the verdict per workload class — that procurement leads actually need.

June 15, 2026Samuel A.13m

Strategy

Establishing an AI-Augmented BPO Operation: The Setup Playbook

Most companies that want to own an AI-augmented BPO operation know the destination but not the route. This is the 90-day setup playbook we run end to end — talent funnel, regulatory wiring, Connect plus Bedrock substrate, training programme, eval gate — stated explicitly so a prospective partner can see how we would build it with them.

June 14, 2026Samuel A.15m

Engineering

When Your Agent Needs to Run Code: Three Real Paths, the runc Reckoning, and the Trade-Offs Nobody Surfaces

A viral LinkedIn post argued that running agent-generated code inside a standard Docker container is a host-breach waiting to happen. Three weeks earlier, three runc CVEs were disclosed that proved the point in concrete. The architectural shift toward hardware-virtualised microVMs is real, the LangSmith Sandboxes release is real, and the engineering trade-offs the marketing leaves out are real too. Four paths, when each one fits, and the path most teams should actually be on.

June 13, 2026Samuel A.14m

Compliance

AI Bill of Materials (AIBOM): SBOM for the Model, the Prompt, the RAG Corpus, and the Fine-Tune Adapters

SBOM became a regulatory artefact three years after Log4Shell, and AIBOM is on the same arc. The substrate is more fluid — model versions, fine-tune adapters, RAG corpus snapshots, prompt revisions, and inference-time configuration all change faster than any release cycle. This piece sets out the seven elements an AIBOM actually contains, where each is generated, and the reference architecture that survives an EU AI Act Article 11 inspection.

June 13, 2026Samuel A.16m

Engineering

Hardened Containers for AI Workloads: The Tier Decision Most Teams Get Wrong

Every AI platform engineer eventually has to answer the question — runc, gVisor, Firecracker, Kata, or something else? The answers in vendor documentation reduce to either 'use ours' or 'it depends.' Both are unhelpful. This piece walks through the four production tiers honestly, the three workload axes that determine which tier fits, the two questions that change the answer, and the hybrid pattern most regulated AI platforms actually ship. With cost, latency, and isolation numbers from the production deployments we run.

June 13, 2026cmdev11m

Engineering

Frontier-Model Lock-In Is Real: How to Architect for Provider Independence Without the Multi-Cloud Tax

The frontier-model vendors and their analyst proxies will tell you the OpenAI-compatible API has solved portability. It has not. Lock-in is real, it lives at four specific layers — tool-call format, prompt investment, provider-specific features, and eval calibration — and the architectural pattern that bounds it is well understood. Claude-first is still the right default for most enterprise reasoning workloads. The question is how much lock-in you accept, where you accept it, and how cheaply you can leave when leverage matters. The honest engineering map.

June 12, 2026Mayowa A.15m

Engineering

The Self-Improving Agent: What SIA Does, What It's Missing, and the Production Pattern That Actually Ships

Hexo Labs open-sourced SIA last month — a self-improving agent that edits both its own scaffolding and its model weights between runs. The benchmark numbers are real and the research contribution is substantial. The walkthrough also has no rollback mechanism, no held-out eval set, no regression guard, and the authors themselves flag the Goodhart risk. We read the paper, the repo, and the walkthrough. Here is what SIA actually does, what the production teams considering it need to add, and the practical pattern any enterprise should actually borrow from this work.

June 12, 2026cmdev12m

Engineering

Vulnerability Response for Self-Improving Agents: Patching a System That Patches Itself

A CVE drops in a runtime your self-improving agent platform depends on. The agent is mid-iteration, holding state, generating its own code, and scheduled for the next safety-eval gate in four hours. The traditional vulnerability-response playbook — patch, restart, validate — has three steps that all fight the agent's autonomous loop. This is the response pattern that actually works on SIA platforms, the kill-switch design that survives a partial patch, and the audit trail that lets a CISO defend the decision in retrospect.

June 12, 2026Samuel A.11m

Compliance

EU AI Act Enforcement: What Actually Changes on 2 August 2026, and What Most Compliance Programmes Have Wrong

On 2 August 2026 the EU AI Act's high-risk-system provisions become enforceable. The compliance industry has produced a small library of frameworks, gap assessments, and readiness checklists. Most of them are reading the wrong sections of the regulation. This piece walks through what actually changes on the day, the four obligations that will produce the first enforcement actions, how Article 14 human oversight is being misread, and the audit-trail pattern that survives a national-authority inspection. Written for the security and compliance leaders who will be on the inspection-receiving end.

June 11, 2026Samuel A.14m

Engineering

Knowledge Base Hygiene for Contact Center AI: What Connect + Bedrock Assumes You Already Did

Amazon Connect AI Agent and Bedrock Knowledge Bases ship a polished pipeline that assumes the corpus on the other end is clean. In practice it is five to ten years of overlapping policy PDFs, agent macros, internal wikis, and contradictory versions. This is the hygiene playbook — the audit, conversational chunking, the conflicting-policy fix, chunk-level sensitivity, the freshness pipeline, and the cost truth nobody puts on the slide.

June 11, 2026cmdev14m

AI Security

AISPM vs AIWPM vs DSPM vs CSPM: The Acronym Map Vendors Don't Want You to See

Four acronyms now compete for the same security budget — CSPM, DSPM, AIWPM, and AISPM — and every CSPM vendor is rebranding into the AI ones without changing the product underneath. This piece names the actual control coverage of each, where the seams sit, and the minimum viable bundle a CISO can defend through procurement and a real incident.

June 10, 2026Samuel A.15m