05 / Field notes

What we build. What we learn.

Security, compliance and open-source AI infrastructure. Practical notes from the team building Xavani and the Security Cloud, written in the open.

Explore open source Read the docs

Read the thinking. Explore the source. Put it to work.

The best AI models for agent work: August 2026

Archive: our 27 August 2026 model comparison. Use the dated snapshot as a starting point for evaluation, not a live price list or a current Xavani default.

27 Aug 2026 · 3 min read · Enternovate

Editorial illustration for “The best AI models for agent work: August 2026”: platform motif in dark indigo and lavender.

Open Source

Why we open-sourced our AI agent gateway

Xavani gives teams a local-first agent they can inspect and adapt. Here is why open code matters, and where configured cloud providers still cross the privacy boundary.

12 May 2026 · 2 minRead

Email Security

Reach DMARC p=reject without breaking email

A staged route from monitoring to enforcement reduces avoidable mail disruption. Inventory senders, check alignment and review reports before changing your domain policy.

28 Apr 2026 · 2 minRead

Compliance

POPIA in 90 days: a practical playbook for SA SMEs

Use 90 days as a planning horizon, not a compliance guarantee. Map personal information, assign responsibility and build evidence with Gavaza and an accountable owner.

10 Apr 2026 · 2 minRead

Offensive

Attack-surface management as code

Make authorised scope, discovery and change review repeatable. Version the asset inventory and turn relevant changes into owned findings instead of another unattended report.

22 Mar 2026 · 2 minRead

Offensive

Red-teaming your own LLM stack

Test prompt injection, data boundaries and tool permissions in a controlled environment. Web auditing and LLM red-teaming are related disciplines, not interchangeable product claims.

14 Feb 2026 · 2 minRead

Open Source

Kimi K3: open frontier intelligence arrives

Archive: a Kimi-focused evaluation note from August 2026. Check the exact model card and licence before treating open weights, context claims or hosted access as deployment approval.

04 Aug 2026 · 2 minRead

Open Source

Qwen 3: why Alibaba's open models keep winning

Archive: what makes a local model useful in practice. For Qwen deployments, checkpoint licence, quantisation, memory and tool reliability matter more than a family-wide recommendation.

25 Jul 2026 · 2 minRead

Open Source

GLM-5.2: built for long-horizon tasks

Archive: evaluating long-horizon work through Xavani. A dated model headline is not a guarantee of sustained reliability, consumer-hardware fit or current benchmark leadership.

01 Aug 2026 · 2 minRead

Open Source

DeepSeek V4 Flash: fast reasoning for agent work

Archive: a DeepSeek-focused note on speed and reliability. Pin the model build and test cost per completed workflow instead of treating an old Flash label as a current recommendation.

01 Aug 2026 · 2 minRead

Open Source

MiniMax M-3: open multimodal with sparse attention

Archive: a practical view of multimodal-agent evaluation. Verify the selected MiniMax endpoint, supported inputs and licence rather than relying on historical architecture or speed claims.

20 Jul 2026 · 2 minRead

Agentic work: what it is and why it matters

An agent can plan, call tools and check outcomes. Useful business automation starts with one measurable workflow, explicit permissions and a human owner for consequential decisions.

15 Jul 2026 · 2 minRead

What advances agents next: five research directions we are watching

Long-horizon reliability, useful memory, tool governance, evaluation and isolation. Five engineering questions for Xavani and its ecosystem, not a claim of solved agent safety.

08 Jul 2026 · 2 minRead

Open Source

Run AI on your own machine: the local-first stack

Local inference can reduce third-party data exposure. Pair a suitable model runtime with Xavani, then verify hardware fit, network access, permissions and the cost of operating it.

18 Jun 2026 · 2 minRead

Security

South Africa's cybersecurity constraints, stated plainly

Limited budgets, specialist capacity and infrastructure resilience shape security decisions. A practical South African approach combines maintainable tools, clear ownership and tested recovery.

03 Aug 2026 · 2 minRead