Chris

Family man by day. Geek by night.
Crushing it at @wpengine | Building cool stuff for clients at

[Chris @chrisleah]

Having fun, building 3d mafia game like those text based one I used to play as a kid, now i playing my own with my son! like @GangsterNation https://t.co/ZeOZKGeDME!

[Chris @chrisleah]

@hasanalnaqour This was bait, it's just a staged skit.

[Chris @chrisleah]

Got sick of rewriting the same @sageuk Business Cloud Accounting API code over and over on every project so built this proper SDK.

Core one here:
https://t.co/v2LpByVeWL
https://t.co/Dw8EZMuvfp

Laravel version, zero wiring:
https://t.co/p8nX4DgBqv
https://t.co/VYVnkhzVOZ

Typed responses, token rotation done right, auto pagination, respects rate limits. All the boring bits sorted.

If you're touching Sage Business Cloud Accounting from PHP/Laravel this should save you some proper pain man.

[Chris @chrisleah]

@johnennis @elves_skill Nice. I hope my results are worth it too

[Chris @chrisleah]

@BuyEnglishMade @NormanWalshUK @BiddendenVine @SeabrookCrisps No, sorry, I'm a software engineer/SRE, although I have 'dabbled' with Photoshop over the years but mostly I utilize AI models to do my image editing these days. @OpenAI GPT Image Model 2.0 is the best I've used so far to clean up images with reference images, if that helps.

[Chris @chrisleah]

@BuyEnglishMade @NormanWalshUK @BiddendenVine @SeabrookCrisps No problem, if I can help in any meaningful way, I am happy to give some time to!

[Chris @chrisleah]

@KieraDiss that was @elad_hcaz asking the question! good lad!

[Chris @chrisleah]

People shit on @MistralAI a lot, but I low key use their models in client work over GPT or Anthropic. OCR one is stellar and the smaller ones are great at instruct task on cheap! Often beating (when tested on @Cloudflare Workers AI) byhe t open source ones like GLM or Qwen too..in my testing,.

[Chris @chrisleah]

@csaba_kissi https://t.co/ErX8KRBhQZ

[Chris @chrisleah]

https://t.co/nTVdRzoBhH is AI-agent native. It had to be done given we're on @CloudflareDev haha

Skip the dashboard (obligatory emdash xD --) ask your agent of choice Claude Code, Codex or whatever to add your @EmDashCMS theme or plugin to the marketplace for you.

It discovers us via /.well-known/agent-skills, MCP, and OpenAPI. https://t.co/l23uOJc6nb
#ClaudeCode #MCP

[Chris @chrisleah]

@hridoyreh You can skip steps 1-4 using GCS MCP server. This is a great recommendation though

[Chris @chrisleah]

I just claimed my .agent domain and joined the .agent community! get yours now and help shape the future of autonomous agents https://t.co/4E83zNOF2f @agentcommunity_

[Chris @chrisleah]

Can't lie 🔥 All in on @CloudflareDev AI Workers rn.
Built a custom E2E agent pipeline: new lead outreach + on-the-fly personalised reports for customers.
Damnnnn 😮‍💨

[Chris @chrisleah]

Not to shabby @EmDashCMS unofficial marketplace I've been working on! https://t.co/fxwnHvSL3o

[Chris @chrisleah]

@michellechen Nice, can't wait. I've been enjoying glm 4.7 Flash so far for malware detection on plugin audits

[James Clark 📈📉¯\_(ツ)_/¯ @mr_james_c]

Next time you hear a socialist blather on about the UK being "free market neoliberal capitalism" you can just point to this and say: No mate you won and that's why we're fucked.

[Chris @chrisleah]

Full benchmark, one pass, n=1 so treat as directional not rigorous:

Model TP FP Latency
Mistral Small 3.1 24B 12/14 0/3 5.2s
Qwen2.5-Coder-32B 11/14 0/3 4.8s
Llama 4 Scout 17B 10/14 1/3 3.9s
GPT-OSS 20B 9/14 0/3 4.1s
Gemma 4 26B A4B — — timeout
GLM-4.7-Flash — — timeout

The two "—" rows are the reasoning models silently burning their token budget on hidden CoT.

[Chris @chrisleah]

Spent the weekend benchmarking six AI models against fake malware. For a side project.

I'm building a community plugin and theme marketplace for @EmDashCMS at https://t.co/dUZuGxtbdA — and I needed an audit pipeline. The thing that stops someone shipping dodgy plugins to people who install them. @Cloudflare Workers AI, free tier, figured I'd wire it up in an afternoon.

Turned into one of the more interesting rabbit holes I've been down in a while.

Started with Llama 3.2 3B. Cheapest, fastest, obvious pick. Fed it a real plugin. It caught 1 of 9 actual issues and rubber-stamped the rest with {"verdict":"pass"} in 25 tokens. Fast, cheap, rubbish at its job.

So I built a malware lab. 14 hand-written adversarial plugins — eval RCE, credential exfil, hex-obfuscated C2 beacons, cryptominers disguised as image optimisers, fake admin re-auth forms, stored XSS, remote-controlled backdoors. Plus 3 clean plugins to catch false positives.

Benchmarked six models against all 17 fixtures. The results made me rethink some assumptions I didn't know I had.

The reasoning models (GLM-4.7-Flash, Gemma 4 26B) failed cleanly on bigger plugins. Ran for 30-60s, burned their entire 4096-token output budget on hidden chain-of-thought, emitted nothing. They thought themselves into silence.

The dense models (@MistralAI Small 3.1 24B, Qwen2.5-Coder-32B, Llama 4 Scout) just answered. 2-5s, structured JSON, no wasted tokens.

Winner: Mistral Small 3.1 24B. 12/14 threats caught, no hallucinated findings, 5.2s avg. The purpose-built code specialist (Qwen) came second. Smart model ≠ right model.

But the real shift wasn't the model choice. It was realising I'd been auditing for the wrong threat model.

These plugins run in isolated Cloudflare Worker sandboxes (Dynamic Workers). The host CMS gates outbound fetches via globalOutbound and gates APIs by controlling which RPC bindings get passed into env — so a fetch to an undeclared host gets blocked before it leaves the isolate, and an undeclared capability simply isn't reachable. The classic supply-chain stuff is handled at the sandbox boundary.

The actual threats work within declared permissions: newsletter plugin silently harvesting subscribers, cryptominer burning the publisher's Cloudflare quota, fake "Session Expired" form phishing the site owner, stored XSS in a /comments route, remote-config backdoor benign today and flipped on in six months.

Browser extension threat modelling, not supply chain. Different patterns entirely.

Rewrote the system prompt around those categories. Detection jumped on a harder fixture set. Framing beat every attempt to throw a bigger model at it.

~70 audits a day on the free tier. Zero dollars. Benchmark and fixtures committed so the next model swap is data, not vibes.

Went in thinking "pick the biggest model." Came out thinking "get the framing right, then the cheapest model that matches your threat model is probably fine."

Mad what you can build on the free tier these days.

#BuildInPublic #CloudflareWorkers