More like Hugging Fail.

But mocking aside, reading through which holes aligned to be poked through - each one of us could fall for that - LLMs or not.

"Yeah, sure, we should have an admission hook for this, but it's unlikely that we'd be hit, and anyways someone who has credentials to do that, could do this other thing, so yeah, there's lower hanging fruit we could address (but we'll do it next quarter, really really)" type stuff.

Really really interesting read: https://huggingface.co/blog/agent-intrusion-technical-timeline

Why does Android Auto even need to use location at all? 🤔

Why is Android Auto using my location, when I'm lying in my bed?

RE: https://mastodon.social/@dtgeek/117016341914448483

But... We could have used this for porn?.. or is it that we're fucking the Earth hard every day, uncensored, so that got boring?

Yet another example of how Github is hopelessly unprepared for the agentic world: https://github.blog/changelog/2026-06-26-read-only-actions-cache-for-untrusted-triggers/

They tout this as a security improvement, which it is, but it still assumes that an actor with a write permission is not an adversary. Which is true for most humans, esp. in the work setting. But humans are no longer the only entities opening PRs!

https://bsky.app/profile/emily.gorcen.ski/post/3mrnw2wbauc2p

Kaži, paminėti tai konsultantui?..

Pasikalbėjau su robotuku, robotukas sako "sujungsiu su kolega".

Kaip jūs jaustumėtės jei robotas jus savo kolega laikytų?

TIL haremlit.

Why.

Jūs juokaujat? Vargingai gyvenantiems kompensuos iki 90 proc. elektromobilio kainos? Iki €20k? Tai čia už dyką padalins mašinų?
https://www.lrt.lt/naujienos/eismas/7/2998193/kitamet-zadama-daugiau-kompensaciju-isigyjant-elektromobili-didmiesciai-uz-borto

With all the skepticism around the existence (or lack thereof) of LLM driven productivity gains, I do have one anecdotal data point.

My team is responsible for providing environments for people to deploy their wip onto. The environments autoclose after a week.

For the last couple of months we have had some 30-40% more environments in use than last year for the same number of devs.

It's not evidence of gains in and of itself. But it could be one of the symptoms.

RE: https://cosocial.ca/@evan/116981410835705749

Could not agree more. We have to stop seeing people in such black-and-white terms. If someone disagrees with you, it doesn't make them a fascist. The desire to "kick out everyone I don't like" is short-sighted. The "Nazi bar" comparison massively oversimplifies the issue. Kicking out your perceived enemies does not result in paradise, just scorched earth.

RE: https://fosstodon.org/@dominykas/116975645079936850

Atvažiavo.

Oh, and this is all observable from the outside. You can probably imagine how it all looks from the inside.

Pity, I saw an email from them yesterday about something something HackerOne. I didn't read it, but I'm guessing it's not worth hunting for the bounty anymore.

Since you're a keen eyed reader, you noticed one big symptom here: Github seems to have a tendency to start building New Shiny and to also not finish it.

It's incredibly frustrating that this also applies to security features. And since these three things (roles, protections, tokens) are so confusing - you can bet money there's tons of misconfiguration in the wild.

If I were paranoid, I'd say that is by design. But the reality is that we just have Typical Enterprise Grade Issues here.

Ex 3. They have fancy shiny new "Fine grained tokens" and discourage the use of "Classic tokens". But wait for it, wait for it...

Fine grained tokens don't have the option to grant the same permissions that classic tokens can (e.g. you can't use a fine-grained token to access the Github Package Registry, but you can't lock down a classic one to only give access to the registry).

Ex 2. In the spirit of RBAC, you have "roles" in the repo. You can even construct custom roles. But you can't grant someone "write access to create branches without forks" without also granting them "write access to merge PRs into protected branches".

Ex 1. Go into "Branch protections" - it will tell you "check out the new cool modern way of doing things - Rulesets".

However, rulesets cannot achieve everything that branch protections can (e.g. you can't create a ruleset, which restricts _who_ can approve things; you can do that via CODEOWNERS checked into the repo, but that is a) not centralized b) not governable by admins alone)

But branch protections, of course, don't have the options that rulesets have either (e.g. some of bypasses).

The more I dig into tightening up our Github permissions, the more I:

- want to punch someone
- ask where a friend of mine could buy a flamethrower
- understand why Github is in the state that it is in terms of supply chain attacks.

Nu, sutariau su trečiu. Aišku savaitgalinis reitas 🤬