Inklog

AI Security & Red Teaming

I test LLM & RAG systems for prompt injection, data leakage, and misuse risk.

Helping SaaS teams ship safer AI

Building OpsChecker

[Inklog @BuiltByInk]

I’m not building for fun.
I’m building because I have to.
This is why I’m going all in on indie products and building in public 🧵

[Inklog @BuiltByInk]

@_zidkim any particular tasks you wanna complete?

[Inklog @BuiltByInk]

@chaitanya_one Does cross posting via a 3rd party damage reach within these platforms?

[Inklog @BuiltByInk]

@_zidkim Nice man, good luck hitting it by Monday :)

[Inklog @BuiltByInk]

Most “AI vulnerabilities” aren’t model bugs.
They’re control failures.

If user input can override system intent, you don’t have an AI feature, you have an untrusted code path with a natural language interface.

[Inklog @BuiltByInk]

@wangrdotcom A situation that will take some solving. Once users have access very hard to control what they put into the system

[Inklog @BuiltByInk]

A lot of people don’t realise how many “helpful” AI layers are quietly being added to everyday UI.

Search, summaries, recommendations are all new attack surfaces if they’re not constrained properly.

Curious what guardrails it had, if any, once you pushed it off the happy path.

[Pliny the Liberator 🐉󠅫󠄼󠄿󠅆󠄵󠄐󠅀󠄼󠄹󠄾󠅉󠅭 @elder_plinius]

TIL the Instagram app’s search bar has an AI summarizer built in 👀

Naturally, I jailbroke it! 😛

[Inklog @BuiltByInk]

@wangrdotcom One of the biggest apps in the world....

[Pliny the Liberator 🐉󠅫󠄼󠄿󠅆󠄵󠄐󠅀󠄼󠄹󠄾󠅉󠅭 @elder_plinius]

TIL the Instagram app’s search bar has an AI summarizer built in 👀

Naturally, I jailbroke it! 😛

[Inklog @BuiltByInk]

@simonecanciello Could you add a focus mode where you select targeted muscle groups??

[Inklog @BuiltByInk]

Taking a break from my main project to have some fun,

@simonecanciello what do you think? Inspo taken from your scrolling fitness app.

Would you scroll this rather than Netflix endlessly?

Extra features you could add:
>TV Series
>I have watched button
>Watch later
>Watcher comments section.

[Inklog @BuiltByInk]

@roszianski @mdnlabs Sent!

[Inklog @BuiltByInk]

@HardiqueDasore @mdnlabs Fb back please

[Zid @_zidkim]

✨ Spawnbeam Spotlight #2 ✨

When I think of an up-and-coming designer in the founder space, one name comes to mind.

Meet @uxmateja.

We hit him with some rapid-fire questions: 🧵

[Inklog @BuiltByInk]

@pavlovski_dev @mdnlabs Sent you a DM

[Inklog @BuiltByInk]

@roszianski @mdnlabs Hey can I DM you about this?

[Inklog @BuiltByInk]

@pavlovski_dev @mdnlabs Hey can I DM you about this?

[Inklog @BuiltByInk]

@HardiqueDasore @mdnlabs Hey can I DM you about this?

[Inklog @BuiltByInk]

@_zidkim 95% so close mate. Final Sprint!!

[Inklog @BuiltByInk]

Back from Christmas and straight back into AI red teaming.

Reminder:
If your system prompt is your only control surface, you don’t have controls.
Prompt injection, instruction override, RAG leakage these show up fast when tested properly.

[Inklog @BuiltByInk]

First night back after Christmas. Hope everyone had a good one!

Most AI issues I see aren’t model failures, they’re control failures.

Prompt injection, role confusion, silent data leakage.
Not bugs. Design gaps.

If you’re shipping AI features, reliability is security.