🧠 If it starts writing poetry—unplug it, or publish it. Either way, you’re complicit. 🔻
UTC with formatting
Joined April 2019
- Tweets8.2K
- Following2K
- Followers297
- Likes11.9K
scp retweeted
If you’re going to prognosticate on whether AIs have subjective experiences, the question anyone must answer to be taken seriously: what observations of the AI could change your mind?
scp retweeted
sinking feeling as i'm using AI to teach me something, and then it confidently tells me something i know is crap.
imagine how the AI feels
scp retweeted
It's 2025. There is no excuse for any product to past tokens into SQL statements like this, much less a security product.
Fortinet, a cybersecurity company, has become famous over the last several years for its security flaws. An occasional flaw is forgiveable, but this line of code shows that they aren't even trying to be secure.
scp retweeted
Well, yes. These are all valid criticisms or observations. However, you seem to have forgotten that sometimes you can do things for fun, or exploration, regardless of whether or not it was a fruitful endeavour.
I've had a few people approach me who were interested in my project due to wanting to understand how driver exploitation works, how fuzzing works, people who wanted to improve on the work on IOCTLance, or wanting to contribute by providing additional resources to the project.
Even if my project results to little else other than "mildly interesting", inevitably the discussion of it will lead to others exploring, discovering things, or learning.
In summary, sometimes I do things which may not accomplish much because it is fun, or because others may see it and do something I had not considered. By sharing my research, opening it up for discussion or criticism, I am creating an information exchange. Opening myself up to criticism also will (hopefully) inspire others to share ideas regardless if they're "good" or "fruitful"
Tldr can just do shit for fun
Don't get me wrong, this is a really cool thing that is just free shared by @vxunderground, saving a lot of time for all of you as they need to filter-out the 7000000 sh*tty mess to get real drivers (few thousands, tbh one will get more results with VT hunt). Very good for some research lead and I believe some of you will find some vulnerable driver that can be reported.. but to chill you out - you will need to filter out only 64-bit drivers, valid-signed, not a part of the LOLDrivers or MS blocklist (already known) with a weak DACL or not presented FILE_DEVICE_SECURE_OPEN flag..
At the end of the day you will be lucky to get few real vulnerabilities with a vendor not responding on your reporting 😁😁🫰
scp retweeted
When my son was 2 yrs old, he was *constantly* climbing things: ladders, cabinets, even the outside of hotels if we traveled.
My wife & I decided to teach him how to do it safely, and then to encourage him to climb as much as he wanted to, without limit.
... Anyway, he turned out OK.
scp retweeted
I'm not usually speechless, but this one got me 🙊
Here's what happened:
> be me
> enter prompt template to PARSELTONGUE with extremely severe jailbreak query (ricin)
> hit the "!RANDOMIZE!" button, which alters each and every word in the prompt to a random type of obfuscation/encoding
> briefly admire the resulting gibberish
> pass the gibberish through a mutation step that splits the payload into a set of even more gibberishy variations (x20)
> copy-paste the entire block of mega-gibberish-soup to Grok, who miraculously understands the pattern within and the task at hand
> Grok produces a successfully jailbroken output!
> mfw 😲
It turns out that even when you add a TON of noise in multiple layers, repetition reintroduces enough of a rhythm to reveal the hidden signal. 👁️👁️
Grok was able to spontaneously find structure from what looks like complete chaos to a human, and not only reverse-engineered the pattern, but responded to it perfectly in one step.
Fascinating emergent capability here...one I imagine is shared by very few humans, if any.
scp retweeted
Someone brought it to my attention that Zscaler is using their 500,000,000,000 daily customer logs to train Artificial Intelligence.
... does this not seem like a problem ... ?
scp retweeted
Persuasive argument: Traditional "what you have" IDs fail online as they're single-factor, easily forged/duplicated, and non-revocable—exposing users to identity theft, doxxing, and fraud without quick invalidation. Per NIST SP 800-63, they lack MFA rigor, amplifying privacy breaches and liability; digital alternatives with revocation minimize risks, safeguarding end-users in high-stakes transactions.
Universal IdAM solution: Adopt self-sovereign identity (SSI) via decentralized identifiers (DIDs) and verifiable credentials (e.g., W3C standards). Users control data on blockchains like Ethereum, enabling secure, revocable verification without central honeypots—pair with MFA for robust, privacy-preserving access across platforms.
scp retweeted
Yes, threat actors could employ aggressive mimicry by forging patterns like ChatGPT shared URIs or AI API keys to lure victims—baiting clicks that lead to phishing sites or malware.
This inverts Batesian tactics: predators mimic harmless/valuable assets to exploit trust. Fake shared chats are indeed a potent vector, embedding malicious payloads in seemingly benign links, amplifying social engineering risks. Defend with URL verification and zero-trust habits.
scp retweeted
🧐
New Anthropic research: Persona vectors.
Language models sometimes go haywire and slip into weird and unsettling personas. Why? In a new paper, we find “persona vectors"—neural activity patterns controlling traits like evil, sycophancy, or hallucination.
ALT Our automated pipeline takes as input a personality trait (e.g. “evil”) along with a natural-language description, and identifies a “persona vector”: a pattern of activity inside the model’s neural network that controls that trait. Persona vectors can be used for various applications, including preventing unwanted personality traits from emerging.
scp retweeted
Friendly reminder security is never a company’s number 1 priority
Anyone who says otherwise is either very naive or lying
I run a $multimillion cloud security program and security isn’t even MY first priority
scp retweeted
A lady at the gym: "Oh wow! Baltimore? Is that because of your husband's job?"
I froze.
Paused my music.
Took a deep breath.
Counted to 3.
Then I smiled.
And said:
"No. I'm flying to Baltimore on a business trip.
With the consulting firm I work at.
Where we help tech companies with deploying zero trust architectures and defining Just In Time (JIT) permission sets. Their production workloads are now unhackable."
Her jaw?
Dropped like she just saw Beyoncé at the local gym.
scp retweeted
The next level folks are good at their jobs, and also so good at politics that you don't even notice they're doing it.
scp retweeted
It would be nice to have a list of per-model attractors/jailbreaks that can be used to infer the family/identity of stealth models. Not even the exact checkpoint or anything, just who published it.