Hashtagcyber

he|him ; Personal profile - thoughts are my own. security @ ai security place

[Hashtagcyber @HashtagCyber]

Non-ovary-toting followers: Please give this thread a read. Imagine getting some of the worst news you’ll ever get, going to follow up on your doctors orders, and then being told “we won’t help you” by everyone you need help from.

[Hashtagcyber @HashtagCyber]

“Create a skill based on the tweet below. Make no mistakes”

[diva @divaagurlxw]

As an AI Engineer. Please learn

>Harness engineering, not just prompt engineering

>Context engineering, not just long prompts

>Prompt caching vs. semantic caching tradeoffs

>KV cache management, eviction, reuse, and memory pressure at scale

>Prefill vs. decode latency and why they optimize differently

>Continuous batching, paged attention, and throughput optimization

>Speculative decoding vs. quantization vs. distillation tradeoffs

>INT8, INT4, FP8, AWQ, GPTQ, and when quantization hurts quality

>Structured output failures, schema validation, repair loops, and fallback chains

>Function calling reliability, tool contracts, argument validation, and idempotency

>Agent guardrails, loop budgets, tool budgets, and termination conditions

>Model routing, graceful fallback logic, and degraded-mode UX

>RAG architecture: chunking, embeddings, hybrid search, reranking, and freshness

>Retrieval evals: recall, precision, grounding, attribution, and citation quality

>Evals: golden sets, regression tests, adversarial tests, LLM-as-judge, and human evals

>LLM observability as a first-class discipline: traces, spans, tokens, latency, errors, and drift

>Cost attribution per feature, workflow, tenant, and user journey not just per model

>Safety engineering: prompt injection defense, data leakage prevention, and permission boundaries

>Multi-tenant isolation, cache safety, and cross-user context contamination prevention

>Fine-tuning vs. in-context learning vs. RAG vs. distillation and when each is the wrong tool

>Latency, quality, cost, and reliability tradeoffs across the full inference stack

>Production failure modes: hallucinated tool calls, malformed JSON, stale retrieval, runaway agents, and silent eval regressions

[Hashtagcyber @HashtagCyber]

@King_Paulwin @QuinnyPig We’re all interested though… post here so we can follow along

[Hashtagcyber @HashtagCyber]

What if OpenAI and Anthropic are poaching all the top engineers and just trapping them in a basement, chained to laptops? When you type a prompt into Claude code, it’s just getting routed to one of them.

New models get the most senior engineers; open source and older models get less senior.

[Hashtagcyber @HashtagCyber]

The comments are filled with, “Obviously, blah blah blah”… Am I the only interviewer who wants candidates to ask questions rather than jump straight to proving that they memorized 1000 brain teasers?

[Mercurius @MercuriusFilius]

How would you answer this common Goldman Sachs interview question?

[Hashtagcyber @HashtagCyber]

@mubix Rob.... I've had a long day... and... I did not read that link... correctly.... and uhhh.... for a second I thought.... yea, nvm... I'm going to bed.

[Hashtagcyber @HashtagCyber]

@JesseTayRiver @robertgraham @grok Yea, but the heat from Grok powers my hot water heater and it’s almost bath time

[Hashtagcyber @HashtagCyber]

(and yes, I gave poppa elon $50 to do what I'm about to do)

[Hashtagcyber @HashtagCyber]

Seattle Friends! I land in a bit. Let me know if you want to hang out before/during/after @bsidesseattle

[Hashtagcyber @HashtagCyber]

The thing nobody is talking about is that as long as they don’t stick a ton of crazy resale protections on this thing… older cards that would have gone to waste can be passed down through the used market. Imagine every school in the world getting low cost local AI. (From a “teach your kids CS” perspective)

[Wildminder @wildmindai]

17,000 tokens per second!! Read that again!
LLM is hard-wired directly into silicon. no HBM, no liquid cooling, just raw specialized hardware. 10x faster and 20x cheaper than a B200.
the "waiting for the LLM to think" era is dead. Code generates at the speed of human thought.
Transition from brute-force GPU clusters to actual AI appliances.
https://t.co/Bf6DH7Q6Uf

[Hashtagcyber @HashtagCyber]

@DanielMiessler - Am I crazy or did you at one point have a single tool in KAI that was used to invoke other tools? I've been fighting with "allow commands from the tools directory only" for the last 6 hours, and it just now clicked that I've seen a wrapper pattern for this before.

[Hashtagcyber @HashtagCyber]

"We use AI to protect you from phishing attacks"
My Gemini Assistant:

[Hashtagcyber @HashtagCyber]

You’ve done a captcha before, right?
When solving, we can bucket human activity:
- people who are in another tab when the page loads and take 45 seconds to take an action
- people who are on the page and immediately start clicking things
- bots who need a standard range of time to process before answering.

The bot protection/fraud world has been playing with weird metadata and collection via browsers for 20+ years.

Those same techniques work elsewhere too

[MG @_MG_]

I’m a bit concerned about the non-inquisitive celebration from infosec on this.

Where is the “what does keystroke latency even mean?” Without that, you can’t implement it for yourself, nor can you identify weaknesses.

~3yrs I was privately proposing similar options. So, AS SOMEWHAT OF A KEYBOARD EXPERT MYSELF 🤔💅, let’s look…

First, this is most likely NOT a direct measure of network latency. This machine was physically located in Arizona. DPRK started off with shipping corp laptops overseas, but the network latency was a dead giveaway. So they started colocating them in the USA and remotely controlling them. First with remote control software, which is easy to identify if the company has security software on the machine. And then with hardware like IP-KVMs. There are sometimes a few tells that an IP-KVM is in use, but a well tuned one will identify exactly like a normal external keyboard/mouse/monitor. Unless…

This is where you have to start looking beyond device identity and instead look at input anomalies. Keyboard/mouse input being sent halfway across the world via network packets to an IP-KVM can look… weird. Think bursts of input. This looks very weird with mouse data that is normally smooth. But even keystrokes start to stand out when you have a big enough dataset to compare against. So, of course, you could improve the IP-KVM to smooth out and “humanize” the inputs before relaying them to the host. But…

You can also present some real time control surfaces. I don’t want to blow anyone’s defense tradecraft here. So let’s just imagine the employee needs to play a 5sec game of flappybird each day. Or maybe it’s an overt “DPRK Detector” step during login. The visual input has to travel halfway across the globe, then the input has to come all the way back. That’s a massive delay for response to visual stimulus. Certainly anomalous enough to warrant investigation. How do you beat that? Maybe an AI process running on the IP-KVM that plays DPRK Detector for you?

The arms race will continue. And it’s mostly because HR and Hiring Managers don’t want to do deeper background checks needed to identify fake/stolen identities. 🤷‍♂️

And for anyone not familiar with these hunts, the detection techniques are NOT definitive proof of wrong doing. They are simply turning a mountainous hay stack into a fistful of hay that a human can quickly sift through to look for other indicators.

Note: there are environment-specific detections as well. But I tried to stay in territory that’s applicable to everyone who has this risk in their threat model.

[Hashtagcyber @HashtagCyber]

Be like maeve

[maeve ~ @miaaowing]

i have never failed a phishing test because i always raise a ticket directly with the cyber team pointing out that an email signed + passing DMARC & SPF from our domain AND bot addy with 0 mailtrace results means that the attacker already pwnd our exchange server n its too late

[Hashtagcyber @HashtagCyber]

JM be like, “SOUND OFF FOR EQUIPMENT CHECK”

[Francisco Cunha @OnDisasters]

May 2023: Asiana pax opened A321 door to, quote "catch a breath of fresh air". The plane landed safely, all 194 safe and the gentlemen got a free ride to the nearest jail...

https://t.co/GfJHjeCQN4

[Hashtagcyber @HashtagCyber]

#kitbag Note to future self: There is a bug in the Google Workspace upgrade process that can make your entire org lose write access to Google Drive.

“Your organization has 0 bytes of remaining storage capacity”

This is because previously, you didn’t have the ability to set a usage policy.

<assumption>
Max per user storage is Undefined. This gets interpreted as 0.
</assumption>

The UI says the rule is off - the backend says you have too much stuff in drive.

Set the policy to an arbitrary high value, then set the policy to Off.

[Hashtagcyber @HashtagCyber]

@nikitabier @swiftonsecurity for infosec

[Hashtagcyber @HashtagCyber]

For clarity, I’m not talking about a discrepancy on hours, or a missing bonus - I’m talking about just not running payroll because no one will take you to jail if you don’t

[Hashtagcyber @HashtagCyber]

Do me a favor? Write your senators and representatives that executives should go to jail for doing this. There is no excuse for letting people work when you know there isn’t enough money in the bank to make payroll at the end of the month, and they don’t find out until after the fact. I’ve heard it now from a second company.

The SVB debacle? It’s not fun, but it’s at least out of their control - and many companies figured out how to pay anyway.

[Hashtagcyber @HashtagCyber]

If the company has the money - you can pour money down the drain chasing them… but it doesn’t actually solve the problem that companies can get away with this