Dustin Hillard

Are you in need of justification for investment in certificate lifecycle management? Tim Callan and I map the crucial role of CLM to NIST's Cybersecurity Framework 2.0. In glorious video.

12

1 Comment

An AI agent tried to scam me today. The script, the response quality, the feel of the conversation, all of these are not there yet, but I'm impressed by the evolution. It's even better at this than poor people in third world countries by not making grammar mistakes and actually approaching you in your native language. For all the good AI automation is doing in the world, in the end it's just a tool and there's always someone who will use it for nefarious purposes. I feel the role of nefarious actors databases like https://lnkd.in/d2QUGi-8 or https://lnkd.in/dCERpTUF will become more important as the AI tooling keeps evolving and improving. The examples I've shared are for URLs, do you know any such databases for checking phone numbers or Whatsapp/Telegram accounts? My concerns are especially with people in desperate situations who are more eager to jump at any apparent opportunity and thus are more vulnerable to these scam attempts. Watch your back out there!

2

Every foundational LLM out there (openai, gemini, anthropic, etc.) is yard sticking each other on "context window" length. "The bigger the better" they say, but that's not always the case -> There's an interesting phenomenon called the Context Window Paradox where once you reach a content window size, output accuracy of the system can actually go down. If you don't know what a "context windows" is for LLMs or want to learn more about the Context Window Paradox I did a short explainer here! https://lnkd.in/ePEbvxTi

22

6 Comments

The harsh reality of CA politics -- the state government works for Big Tech, not the citizens. Some good quotes here. Newsom's stated goal about workable "guardrails" is nonsensical as it's currently physically impossible. That's a bit like saying they are pursuing a way to apply DDT safely to agriculture.... no such method existed. Newsom and CA have zero cred on anything relating to responsible or safe AI. The question is whether that extends to Harris. I honestly don't know, but her history in defending SV and conflicting relationships that include her brother-in-law suggests that it would be difficult for her to regulate Big Tech objectively. Trump is also a wild card. He's so unpredictable that one day he might attempt to order Zuch arrested and the next brag about the power of American Big Tech. The next four years will be interesting. I have much more confidence in SCOTUS.

6

5 Comments

Here we go again. There's another Foxit Reader crash detected by EXPMON Public (https://pub.expmon.com). Unlike the previous ones that were actually patched in the latest version (https://lnkd.in/g_wPzfkJ), this one does cause a crash on the latest Foxit Reader. However, my brief analysis has found that: 1. This is not a malicious exploit, just a real-world PDF sample happens to trigger a bug in Foxit Reader. 2. It seems to be a null pointer dereference crash (so probably not exploitable). Nonetheless, I've sent the sample to the vendor (Foxit Software), for safety reasons. Will share out the particular sample after the vendor finishes their investigation. As reported, recently EXPMON Public has detected some "zero-day" PDF crashes including one notable crafted/exploitable Adobe Reader crash (https://lnkd.in/ghQYYzXR), when it's just on its mission to analyze a (huge) public PDF sample set - WILD. EXPMON is designed and built for detecting any exploitation-related activities. Because application crash is a big and important indicator for potential exploitation, EXPMON carefully reports them. So, if you see Indicator related to application crash when using EXPMON, you should pay attention to that submission. #expmon #ExploitDetetcionInnovations #foxitreader #pdf

24

Competition is a good thing as it forces the market to improve products, services, and price points. Even though SB 1047 was vetoed it has forced the conversation to come up with AI legislation that makes sense that will allow for innovation with privacy and safety. It will be interesting to see the next round of edits to this bill as well as future bills in other states. https://lnkd.in/dN7zfs4D #GenAI #Privacy #California #AISafety #SB1047

38

We've hit peak LLM, and LLMs are underdelivering on their overhyped promises. Sorry to disappoint, but no, GPT-5 isn't going to be AGI or exponentially better than GPT-4. This isn't a secret to the companies developing LLMs, which is why we'll see a shift in focus because when you can’t go wide, you go deep. We need to start asking tough questions, like what exactly are we getting from all of this, and are the tradeoffs worth it? #AI #Privacy #Safety https://lnkd.in/essfv9BE

35

10 Comments

📢 JUST IN: CA Governor vetoed the CA AI Safety Bill S.B. 1047, asking for revisions for the next session! More info here: https://lnkd.in/eVNsZBpx "Mr. Newsom said that the bill was flawed because it focused too much on regulating the biggest A.I. systems, known as frontier models, without considering potential risks and harms from the technology. He said that legislators should go back to rewrite it for the next session." 📚 The AI Policy Newsletter: https://lnkd.in/eS8bHrvG 👩‍💻 The AI Policy Course: https://lnkd.in/e3rur4ff #AIpolicy #ArtificialIntelligence #TechPolicy #AIGovernance

47

1 Comment

SB 1047 is a pending California law that is essentially outlawing large open source models. If you like open source (and I do!) this would be very bad. And now it is reported that one of the advocates of the bill is trying to cash in on it 🤦‍♂️ The bill technically doesn't outlaw large open source models, but it requires "PWC type audits" costing "on the order of hundreds of thousands of dollars to single digit millions." As most open source authors or university research labs don't have that much money, it achieves the same outcome. It would also make AI start-ups vastly more difficult and expensive. The AI community (or at least the part of the AI community that likes open source) unsurprisingly are trying to stop it. Today, the article below reports that "Dan Hendrycks, an executive at Center for AI Safety — the firm whose lobbying arm helped draft the bill — co-founded a company called Gray Swan that offers AI safety compliance tools that seem positioned to provide the type of auditing data the bill would require". https://lnkd.in/gdutbeKq

64

14 Comments

The cost of training top models is increasing 2-3X per year, and at this rate will reach $1 billion per model by 2027. The findings are from Epoch AI and follow their recent analysis of increases in compute for foundation models running at 4-5X per year. They've also provided some interesting data on the relative breakdown of cost components like staff, hardware, and energy. A few thoughts after reading this: - I would assume staff costs won't increase as much as hardware and energy costs, so the relative contribution of human labor to future AI models will decrease, in yet another sign of capital's increasing dominance over labor in the industry (just look at NVIDIA's ever-rising stock price). - If compute is increasing 4-5X per year but cost is increasing 2-3X, is it safe to assume that training runs are getting more efficient? - Some people, like Anthropic CEO Dario Amodei, have said we'll see $1 billion training runs in the next year. Does this mean the 2-3X pace is about to increase? That Epoch is wrong in its estimate? Or that Amodei is overstating the cost, possibly to put fear into competitors who can't keep pace? - Only a few companies are going to be able to train future models. We're likely going to see further consolidation. Perhaps we get to a future with only two large general purpose models, like the Windows/MacOS and Android/iOS duopolies we've seen in the past. Article here: https://lnkd.in/g5SejXen

7

#MyWriting "IP Theft" is a heavy charge being thrown around with respect to large language model (LMM) training. It's a dubious charge based on misunderstanding of what the LLM does. I propose a hypothetical test. The IP Theft charge is three part: (1) LLM are training with specific copyright protected data (TRUE), (2) without licensing (TRUE), and (3) gaining some advantage by doing so (FALSE). We don't have a mathematically sound subtract operation for content on LLMs. So you're going to have to do this in your head -OR- perform on a small base model without the piece of content and assume the effect scales up less. Start with that base model that does not have the particular content in its training set. Ask it a to perform some question / answer or content creative tasks related to the content. For example, Tom Clancy novels. Omit the actual novels from training corpus for base model. Ask the model to write you a short story in the style of Tom Clancy where Eeyore (now public domain) joins Jack Ryan's team and wrecks an op with his gloominess. Even do purely adversarial questions and tasks to try to prove infringement. Run multiple times to account for randomness. Fine tune the base model using just the specific copyright material in question. Re-ask the questions, redo the creative tasks. Qualitatively, did you see a difference? If yes, there might be a case that training on the particular IP delivered an advantage in the LLM. If no, then you can assume the content is drowned out by the average more in larger models. I'm going to shortcut this for you. You will need a very small model and very low lossy compression to see a difference. It just takes too much source material to get the weight balances right enough for general language for any one piece to make a difference. So where does all the "knowledge" about Jack Ryan come from then, if not from the actual novels. Fan forums, reviews, scholarship, etc. He's a popular character. If you own the rights to Tom Clancy's work, your beef is with the audience, not the aggregator. I would love to be able to perform this experiment with real models and data. I would love to publish results, whatever they are. I suspect they would settle the legal debate quickly or move it to technicalities like DMCA and when actual copies are made and kept in the training process. I also think it would demystify the "emergent properties" that many claim or fear these systems have. In that sense, the test hurts the dramatic claims of both sides about the same. While this is about text systems, I suspect similar applies to large enough image systems. If you are a proponent of the IP Theft narrative, would you accept results of such a test as I propose above?

1

27 Comments

When Evan and I started RunReveal, we set out with the goal to make sure no breach goes undetected. A big part of that is making good detection easy. Today, we're excited to announce something that we've been looking forward to for a while: releasing Correlated Alerting! For a long time I felt that security would be easier if we could just make it easier for folks to see what's happening in their networks. While that's true, we found it's only half the equation. What's really important is to be able to make use of that data effectively, something existing SIEMs and other security tools miss the mark on. That's why we built Correlated Alerting, a method to separate your signals from your alerts, where alerts are a compilation of the signals representing a pattern of malicious behavior distinct from business as usual. Imagine you're an investigator on a case. You find clues and evidence suggesting suspicious activity, but in order to make a case to prove what happened you have to correlate the facts into a coherent story. With Correlated Alerting, the signals are the individual pieces of evidence and the alerts represent the collection of evidence that tells the story. We believe that by making it simple to build these kinds of alerts is going to drastically improve the way that security engineers implement detection and response. As always, we're building in public. Learn more about how and why we built it on the blog post: https://lnkd.in/gzuQ-5Qw What do you think? How does your organization think about distilling a stream of distinct events into actionable alerts? If this sounds like the right approach to detection and response, please share with your colleagues! If you think this is a terrible or fraught approach, please let us know why and share with your security friends to see what they think!

15

2 Comments

Every Way To Get Structured Output From LLMs “This post will be interesting to you if: -you're trying to get structured output from an LLM, -you've tried response_format: "json" and function calling and been disappointed by the results, -you're tired of stacking regex on regex on regex to extract JSON from an LLM, -you're trying to figure out what your options are. Everyone using LLMs in production runs into this problem sooner or later: what we really want is a magical black box that returns JSON in exactly the format we want. Unfortunately, LLMs return English, not JSON, and it turns out that converting English to JSON is kinda hard. Here's every framework we could find that solves this problem, and how they compare.” https://lnkd.in/dXvjz8tS

8

The EXPMON Public has reported several Foxit Reader crashes when processing a public PDF sample set. https://lnkd.in/gidnAhMU https://lnkd.in/gCeijxFz https://lnkd.in/grJTxu7b I've manually analyzed them and confirmed they're not zero-day attacks, they're not FPs either. They're even not new Foxit Reader vulns/bugs, as far as I've seen. The crashes could be reproduced in the specific tested Foxit Reader version 2023.2.0.21408 (so not FPs), but not in the latest version. So I'd guess it's a vuln/bug in Foxit Reader just got patched in the last several months. Yeah, sometimes, @EXPMON_ could be used to discover software vulnerabilities (for free), too. You never know what you'd encounter when processing real-world samples! :)

9

1 Comment

inception llms 1/n: I'm currently exploring a new LLM design pattern that I called called "inception llms." This innovative agentic approach involves LLMs like Gemini dreaming up code, which in turn utilizes LLMs to execute particular tasks. Here, I wanted to distill a number of principles useful with Inception 1 - simplicity is the ultimate sophistication: craft a simple prompt to achieve one task only, and do it well. don't over-engineer tasks but keep things as simple as possible. More to follow

28

3 Comments

🚨 Is California’s AI Regulation Going Too Far? SB 1047 is making waves across Silicon Valley, and the entire tech world is paying attention. While it's clear we need safety regulations to keep AI from causing harm, is this bill overdoing it? 🤔 The big tech companies are raising alarms about the potential for increased legal liabilities ⚖️ and how these regulations could stifle innovation 🛑. They’re concerned that heavy-handed rules will slow down progress and put the U.S. behind in the global AI race. And it’s not just the tech giants feeling the squeeze—smaller startups are also on edge. For them, vague testing standards are like an added weight that could make it harder to innovate and compete. 🚀 Of course, there’s the other side. AI safety is no joke. I’m all for responsible AI, but regulations like SB 1047 seem like a step too far. Are we at risk of over-regulating AI in a way that might protect us now but choke creativity and progress in the future? What do you think? Should we push for more balanced regulation that keeps AI development alive and well while still protecting people? Drop your thoughts in the comments!👇 #AI #Innovation #SB1047 #TechRegulation #SiliconValley #AIsafety #AIregulation #CaliforniaAI #ResponsibleAI #AIethics #TechInnovation #AIoversight #AIrisk

14

4 Comments

The acquisition may already have happened by the time this posts, but here's the "why" of the downfall for Lacework: good people left. "Six months after the monster Series D funding round, Lacework laid off 20% of its staff in a bid to strengthen the company's balance sheet. Then in summer 2022, President and Chief Revenue Officer Andy Byron - who spearheaded all of Lacework's go-to-market functions since 2019 - departed. Lacework's CRO responsibilities on an interim basis went to David Hatfield, who joined the company as CEO in February 2021 and became co-CEO in August 2021 after the addition of Facebook engineering head Jay Parikh. But in October 2022, Hatfield left the company, with Parikh telling staff, "Now is the time to unify the company under one leader." Parikh remains the sole CEO of Lacework (see: Lacework Co-CEO David Hatfield Out 4 Months After Layoffs). Parikh spoke with ISMG in February 2023 about Lacework's new attack path analysis tool, which helps organizations understand the havoc specific threats could wreak within their cloud infrastructure. "Our data-driven approach here really drives efficiency and effectiveness for these teams so they can prioritize what they should focus their time on to understand and remediate different areas of risk" (see: Lacework CEO on How to Prioritize Cloud Infrastructure Risks). But these investments failed to prevent continued brain drain at Lacework. After holding headcount steady between 975 and 1,025 employees between September 2022 to August 2023, the company's workforce has shrunk by nearly 24% over the past eight months despite never formally disclosing any layoffs." Finding and retaining good people needs to be job number 1. If it's not, buckle up.

1

For folks still wondering what the #ai hype is all about: the idea that a #GPT decoder model *only* trained to predict the next word can take a plain english* prompt - that expresses an *arbitrary* declarative program and its input data - and produce an appropriate response across innumerable tasks, is unintuitive and nothing short of a miracle in computing. This *is* hype worthy, at least as a technical marvel, and perhaps the answer to the question - why did #ChatGPT get to 100M users faster than any app known to mankind. #llm #openai (*or another language)

177

15 Comments