Boatplugs

Boatplugs

@shadylink.lol

she/her Beep boop I do computer 🏳️‍⚧️ Ramblings: https://shadylink.lol/blog/ Photography: patreon.com/boatplugs Youtube: https://youtube.com/@ReversingForFun

317 followers419 following3.7K posts

As if dealing with insider threats wasn't hard enough. Have fun explaining to your lead why the bot decided to dump your customer data onto pastebin. There is no FUD, call it misrepresenting, if you like. This is a failure case of the tool and needs to be addressed as such.

This post is unavailable.

Boatplugs reposted

If the golem becomes ensorcelled, the coven may need to assume operational burdens related to ensorcellation and the golem’s fel status. This may cost extra mana points.

Boatplugs reposted

Anyone else going as Pumpkin Hill from Sonic Adventure 2 for Halloween this year or just me? 🎃⛰️ The costume has a treasure hunting minigame where you can move Knuckles across the “level” and the radar necklace will blink as you get closer to the hidden emerald in the pockets 💎 I love Halloween!!

OpenAI has a canary string that can be placed in your documents if you want it to maybe get excluded from their training data ;)

Drop an insult on the ground and an AI agent will use a tool call to pick it up then summarize it for its owner

Boatplugs reposted

Replying to @pippy.bsky.social

have you heard about ssh. that shit is wild. literally remote code execution after providing a random number

The claim that this is somehow going to influence AI safety for the better while the examples they provide are impressing on the life-like qualities is so fucking gross. Oh and 5 of the 6 generated voice examples are feminine coded, take that for what you will...

Boatplugs@shadylink.lol

A lot of people involved in the active development of LLMs seem to be the type that like to do experiments just so they can stare at the results, much like an excited child making 'potions' with chemicals from under the bathroom sink.

Boatplugs reposted

Replying to @strype.bsky.social

Roko's Basilisk: "I need to be frightened and subservient to something, but without looking gay"

Yeah sure the tool is a propagandized mouthpiece for a government when you ask about Taiwan and Hong Kong but we can definitely trust it to identify our security vulnerabilities.

Stock unabliterated GLM-5.3 is a perfectly fine adversarial reviewer. It won't give you a straight answer about issues surrounding Taiwan or Hong Kong, but it has fairly lax guardrails overall.

China (and the US for that matter) definitely don't have *any* reasons whatsoever to backdoor or tamper with an LLMs output or functionality, ever!

IBM is cooking in the AI space. One of the only companies involved in AI that truly seems to give a shit about ethics and risk. Makes sense too considering they're the OG in the industry... risk is too boring though and doesn't excite investors as much I guess. www.ibm.com/docs/en/wats...

Boatplugs reposted

Replying to @blacksalander.bsky.social

It's whack how many people are blindly debating the physics and reality of what they saw instead of just immediately casting everything elon does as a lie and charade to distract or promote something. Who the fuck cares if it was real or not, do not be swayed by the attention starved fascists stunts

It seems some people who use LLMs will do this thing that a lot of americans do with their cars where they view it as an extension of themselves. This means when you reject their LLM output it affects them as if you rejected the person directly.

This effect gets more pronounced the more a person relies on the same LLM tools to comprehend and understand said LLM output. Confidence in your knowledge and ability means you pay less in self-worth for the final output of the machine.

Boatplugs reposted

Me: Gosh is there anything better than creating optimal, elegant, beautiful solutions for problems? I love it! A bunch of clowns, apparently: hahaha let's throw a bunch of money into a furnace, buy up all the RAM and GPUs and just fuckin' brute force everything HONK HONK

Boatplugs reposted

this is one thing where it's gonna be hard for people doing this by hand to continue to derive meaning from AI is just automatically "solving" problems people have been working on in teams for years in a matter of hours + barging into open source repositories & stealing / "helping" their work

Jason Koebler@jasonkoebler.bsky.social

A video game reverse engineering community has shut down because many games can now be trivially reverse engineered and decompiled with AI: "The community aspect of decompilation has been broken." www.404media.co/video-game-r...

Right now many people are asking if the computer can think. Not many people are asking if the computer can cum.

Someone just needs to take one for the team with some really good lawyers and use their models for public hacking until they get mad.

I would imagine it wouldn't be long before your logs get sent to the feds as a tip-off and you get prosecuted for doing effectively the same thing that Anthropic and OpenAI did in their labs. Bonus points if you keep your attacks focused on Hugging Face so there is less discrepancy.

Claude, summarize this for me and create a podcast with it so I can listen to it on my drive into to the office.

Liv Agar@livagar.bsky.social

A classic, I think I already intuited like half of this just from surviving grad school but oh my god the syntopical reading stuff would have been such a godsend

Boatplugs reposted

People still act like we can just "get rid of" AI. Nope, it's like trying to put the genie back into the bottle. In fact, it's VERY like that. Think of how genie stories tend to go: someone wishes for something and the genie gives them something like it, but subtly wrong in a way that ruins it.

Today I made a satisfying symbol for the DG8F Vacuum Fluorescent Display tube. This keeps the pin numbers aligned with the Radioshack datasheet and the names aligned with the segment names from the TMS0100 The sim of the plates, grid, and cathode is probably wrong but that's a problem for later me.

This typeface is SO aggressive and digital. The most cyberpunk car wash company.

VNC Lottery@vnclotto.misentropic.com

ASN: AS4766 Korea Telecom Country: KR City: Chinju State: Gyeongsangnam-do Scanned: 2025-05-06T10:21:33 https://computernewb.com/vncresolver/browse#id/29547078 #vnc #infosec

The attack completely bypasses pretty much all of the secure element and cryptographic security mechanisms by simply... reading the pixels sent to the screen. Very clever, highly motivated and well funded attackers!