Victor Junqueira

AI Security | DevSecOps | Blockchain | I break things before attackers do.

[Victor Junqueira @jjunqueiraa]

https://t.co/ccX5quDrUF

[Victor Junqueira @jjunqueiraa]

@mj_0141 That’s really impressive, good job 👏

[Victor Junqueira @jjunqueiraa]

https://t.co/ccX5quDrUF

[Victor Junqueira @jjunqueiraa]

@thechosenberg “I kinda vibecoded the API endpoints” is the most beautiful sentence a bug bounty hunter will ever read

[Victor Junqueira @jjunqueiraa]

@Ze4fer @sluongng Both valid but with Orbstack you’ll run more agents, faster, with less resource waste, and still have excellent practical security

[Victor Junqueira @jjunqueiraa]

@Vl0per thank you! 🙏

[Victor Junqueira @jjunqueiraa]

https://t.co/X4mLDHQD5B

[Victor Junqueira @jjunqueiraa]

@sebastienlorber Solves one symptom. Doesn't kill the root problem:
AI agents have too much shell access by default.

That's the layer PreToolUse hooks fix. https://t.co/FK6tQEW0pi

[Victor Junqueira @jjunqueiraa]

https://t.co/ccX5quDrUF

[Victor Junqueira @jjunqueiraa]

Use hooks so that it doesn't happen to you!
https://t.co/8BOz57XMYa

[𝖊𝖛𝖆𝖉𝖊 @encrypted]

wait... you did what?!

[Victor Junqueira @jjunqueiraa]

Your AI agent has shell access on your machine.
Treat it like an intern with sudo.

Claude Code PreToolUse hooks run BEFORE any tool
executes. Exit code 2 = blocked. Model sees stderr,
stops, asks you.

What 12 lines of Typescript can stop:
> npm/pip install of flagged packages
> reads of ~/.ssh, ~/.aws, .env*
> git push to main
> curl | bash patterns
> DROP TABLE without WHERE
> rm -rf outside /tmp

Almost nobody is using this layer.
I wrote the full breakdown:

https://t.co/yRps6zo5q5

[Victor Junqueira @jjunqueiraa]

@duborges There are many ways to achieve this, but the first and most effective approach is to create a hook that prevents reading or writing into the .env file. I have written more detailed information about this topic here: https://t.co/Q8pdmXBwy9

[Victor Junqueira @jjunqueiraa]

https://t.co/ccX5quDrUF

[Mo @atmoio]

I’ve been trying to find good analogies for this.

The argument seems to be: “AI is intelligent because it does intelligent-like things, therefore it is cope to say it’s not intelligent.”

Some analogies:

- “The moon is really bright. Therefore it is capable of luminance.” Here of course we expose a manner of speaking. The moon reflects the sun’s light. The models reflect our own intelligence. The moon will never be a star.

- “A snail on the bed of a tow truck is really fast. Look, it’s moving from A to B at 60mph, it’s clearly fast.” But of course the snail is borrowing the truck’s velocity.

Notice how there is no controversy in calling the technology large language models because the term is perfectly apt: a map of language. This points to language as constructed by humans as the true source of magic, and LLMs being algorithms that can traverse this map at light speeds.

Before you think I’m being pedantic, understand that the nature of the words we use is precisely what’s at stake. That the moon *looks* bright is incontrovertible. Insisting however that the moon itself has any concept of inherent luminance is when you start to gaslight people into deranged realities that they will not stand for. Attempting to appropriate ageless conceptions like consciousness and intelligence to corporate technology by playing axiomatic word games is insanity.

Large language models do what they do and this is non-controversial. Personifying it with human-like attributes however is totally uncalled for, when it is easy enough for us to define new words that better capture the phenomenon.

I’ve been thinking long and hard about this and I think a good phrase for these technologies can be—hear me out:

“large language models”

[Victor Junqueira @jjunqueiraa]

@atmoio I couldn’t have expressed it better. I’m fed up with the propaganda. During one of my travels, I realized that most people don’t even know what “Claude” or “Anthropic” means, while X is constantly screaming about some impending doom…

[Victor Junqueira @jjunqueiraa]

That’s exactly what PreToolUse hooks are for.

CLAUDE.md is a suggestion the model can ignore. Hooks run deterministically before the tool call fires, so you can block --config.minimumReleaseAge=0 outright and the agent never gets to execute it.

Wrote about locking this down: https://t.co/Njp5BIiBPW

[Victor Junqueira @jjunqueiraa]

@kmcnam1 Everyone saying is “silly stuff” but at the end it’s just money laundering

[Victor Junqueira @jjunqueiraa]

@Maxis_arena I think height matters a lot too

[Victor Junqueira @jjunqueiraa]

@GRITCULT The government is, and they’re savage

[Victor Junqueira @jjunqueiraa]

@thealepalombo Pretty as always

[Victor Junqueira @jjunqueiraa]

@brunopinheiroms E a questão da privacidade dos dados? Você pode desativar nas settings do Claude e se for enterprise é desabilitado por padrão. Mas se usa um modelo Chinês eu tenho minhas dúvidas quanto a isso…

[Victor Junqueira @jjunqueiraa]

@kloss_xyz The difference is that with google you pay straightforward + pay with your data.