Joshua Saxe @joshua_saxe Twitter profile | Pikagi

Pikagi

Joshua Saxe

@joshua_saxe

3,196

Followers

1,173

Following

221

Media

3,044

Statuses

AI+cybersecurity at Meta; past lives in academic history, labor / community organizing, classical/jazz piano, hacking scene

Wichita, KS

https://t.co/mwRl3jPKic

Joined May 2013

Don't wanna be here? Send us removal request.

Pinned Tweet

@joshua_saxe

Joshua Saxe

2 months

With today’s launch of Llama 3.1, we release CyberSecEval 3, a wide-ranging evaluation framework for LLM security used in the development of the models. Additionally, we introduce and improve three LLM security guardrails. Summary in this 🧵, links to paper/github at bottom:

Tweet media one

2

32

117

Last Seen Profiles

@Brighton64bcn

@graemeholmes84

@osmanuzun37

@LoveSumCyd

@DrIvyCo

@Elenmai85

@UweBrandes

@wokcoholic

@ianwalker_ian

@Flowers898219

@jwalker0522

@ProSanitate

@Chocolate_Dady3

@ffffffyy882

@LFC_PAUL_

@EslaEarth

@azmohf

@Bunnyzen21

@emxxle

@DealinTheCards

@papmantan69_

@ibrahym_pharyda

@Aptomingos

@Omerta_FPS

@waru_aber

@_koshurhasa

@UrracaMary66210

@dazzlehousee

@jandakembangstw

@GratignyGators

@bokeplokalmalam

@kahadu

@VitaEcuador

@AI_VISUALS

@p_ribeiro

@i_7md

@joshua_saxe

Joshua Saxe

4 years

How to evaluate a cybersecurity vendor's ML claims even if you don't know much about ML (thread). 1) Ask them why they didn't solely rely on rules/signatures in their system -- why is ML necessary? If they don't have a clear explanation, deduct a point.

21

304

1K

@joshua_saxe

Joshua Saxe

5 years

1\ Surprisingly, you could build a very mediocre PE malware detector with a single PE feature: the PE compile timestamp. In fact, I built a little random forest detector that uses only the timestamp as its feature that gets 62% detection on previously unseen malware at a 1% FPR.

6

109

277

@joshua_saxe

Joshua Saxe

4 years

6) If they claim "zero false positives!" or "zero false negatives!" don't deduct any points, just end the Zoom call.

4

19

267

@joshua_saxe

Joshua Saxe

5 years

1\ I've written a little compiler to ship ML models as standalone Yara rules, and done proof of concept detectors for Macho-O, RTF files, and powershell scripts. So far I have decision trees, random forests, and logistic regression (LR) working.

Tweet media one

13

94

248

@joshua_saxe

Joshua Saxe

2 years

@halvarflake This is like a UX version of Goodhart's law.. if a user-behavioral signal is used to accelerate behavior its utility as a signal begins to degrade.

2

20

212

@joshua_saxe

Joshua Saxe

3 years

I'm committed to not being snarky on Twitter, so I'll just say, the below is a case study in why we need more tech expertise in government.

@GovParsonMO

Governor Mike Parson

3 years

Through a multi-step process, an individual took the records of at least three educators, decoded the HTML source code, and viewed the SSN of those specific educators. We notified the Cole County prosecutor and the Highway Patrol’s Digital Forensic Unit will investigate.

Tweet media one

5K

337

1K

5

23

208

@joshua_saxe

Joshua Saxe

5 months

There's a high BS factor in both AI and cybersecurity, and this WSJ article multiplies them together

Tweet media one

21

23

186

@joshua_saxe

Joshua Saxe

4 months

The vibe of AI policy should be "we don't know what's going to happen, so we need a very tight observe/orient/decide/act loop where we incrementally observe and address AI's societal effects". This in contrast to the vibe of pretending to know how the coming years will play out

9

24

149

@joshua_saxe

Joshua Saxe

5 years

Some infosec knowledge is useful for months (knowledge of a given campaign), other knowledge, for years, (TTPs), other knowledge, for decades (the halting problem). Here's a "Pyramid of Pain" (cc/ @DavidJBianco ) inspired model of knowledge in cyber I find useful for myself.

Tweet media one

3

55

125

@joshua_saxe

Joshua Saxe

4 years

3) Ask them where on Wikipedia you can read more about the approach they took. If you can't read about it on Wikipedia, ask them where their paper is in the peer-review and on arXiv. If the paper doesn't exist / is a "trade secret", deduct 3 points

1

5

122

@joshua_saxe

Joshua Saxe

4 years

2) Ask them how they know their ML system is good. Where does their test data come from? How do they know their test data is anything like real life data? How do they monitor system performance in the field? If their story isn't convincing, deduct a point.

1

4

116

@joshua_saxe

Joshua Saxe

8 months

@Tyler_A_Harper Gather some vague impressions of a diverse and non-monolithic group and then write a poetic, damning critique of them with no evidence

1

0

114

@joshua_saxe

Joshua Saxe

4 years

5) If they're now down a few points, ask them if they'd like to fess up and say they're really mostly just using rules and claiming to use ML to satisfy investors and industry analysts. If they fess up and then describe a solid rules-based approach, give 'em a point or two back.

1

6

111

@joshua_saxe

Joshua Saxe

5 years

1\ Malware sandboxes are useful but extremely limited. Here's a malware call graph, and in red are the functions the malware actually *executed* when run in a sandbox -- a miniscule fraction of the malware's potential badness!

Tweet media one

10

43

103

@joshua_saxe

Joshua Saxe

5 years

Thread on cognitive biases in cybersecurity I've noticed: Maginot Line: you got breached by an impersonation attack, so you go buy an anti-impersonation solution and assume you're much safer. Sort of like checking people's shoes at the airport.

7

35

99

@joshua_saxe

Joshua Saxe

4 years

Infosec friends RT please: As attackers increasingly exploit the health crisis to compromise users, we should be sharing what we're seeing with one another. I'm starting a non-vendor-aligned Slack to this end. Please join and responsibly share intel!

4

108

97

@joshua_saxe

Joshua Saxe

3 months

After a couple days playing with it it seems Claude 3.5 Sonnet unlocks LLM applications that weren't possible with GPT4 and challenges the "LLMs have plateaued" thesis

9

3

93

@joshua_saxe

Joshua Saxe

4 years

4) Ask them why they didn't take a simpler approach than the approach they took. If they can't explain, deduct a point. If they say they tried other, simpler approaches, but can't show comparison data, deduct a point.

1

5

89

@joshua_saxe

Joshua Saxe

5 years

1/ Not everyone in security needs to learn machine learning / mathematical modeling, but getting Bayes rule and ROC curves into common language would elevate the discussion around attack detection a lot, I think. In case it helps, here are references...

1

23

84

@joshua_saxe

Joshua Saxe

4 years

You don't have to know much about ML to use something like the system I'm pitching here. Just don't get intimidated by obfuscation attempts from the vendor (common tactic) and keep pressing for clear and convincing answers.

1

6

81

@joshua_saxe

Joshua Saxe

4 years

7) Ask for examples of attacks the system missed and explanations as to why. Deduct points if they can't think of any misses or can't easily produce a story about their system's known weaknesses.

1

4

80

@joshua_saxe

Joshua Saxe

4 months

I rarely quote-tweet negatively, but this toxic, glib attitude towards working class people's livelihoods should have no place in the AI community

@airkatakana

Air Katakana

4 months

give me a robot and a gpu i could put this guy and everyone he knows and loves out of a job in less than a week

1K

895

26K

2

6

82

@joshua_saxe

Joshua Saxe

5 months

With the Llama 3 launch this morning we launched CyberSecEval2 and CodeShield: CodeShield is a secure coding guardrail system that filters a wide range of insecure coding practices from LLM completions at inference time. 1/x

6

19

80

@joshua_saxe

Joshua Saxe

7 years

@dakami Definitely an astute way to think about complexity. But I think insiders don't see "reality", rather they have their own refined but still lossy abstractions. Nobody really understands complex systems- everyone has a partial view optimized for the specialized role they play.

1

11

75

@joshua_saxe

Joshua Saxe

4 years

8) Now ask for a live demo where you control the inputs and test the system on both malicious and benign data. Give them some points depending on how their system performs on a test that you yourself have designed.

1

4

67

@joshua_saxe

Joshua Saxe

1 year

Making this deck for my Defcon AI Village keynote took an inordinate amount of time because it meant publicly murdering my darlings: the ~80% of MLsec R&D efforts I worked on over ~10 years and which never reached deployment🧵

Tweet media one

3

13

71

@joshua_saxe

Joshua Saxe

5 months

If you and I have the exact same LLM on our local machines we can send one another lossly compressed information that takes advantage of the stunningly low perplexity of modern genAI models can't we? I've heard very little about such applications in practice

24

4

68

@joshua_saxe

Joshua Saxe

3 months

If you're not a doomer and agree to assign a p(doom) you've already in a sense lost the argument. The push-back should be 'what basis do we even have for estimating this quantity?' and 'how is framing the safety discussion eschatologically helpful'?

5

8

65

@joshua_saxe

Joshua Saxe

4 years

Also, what do you do when you have your final point count? Depends on lots of other factors, including the other tech that the vendor is selling with the ML system, and variables like price and the overall value of the technology in your ecosystem.

6

2

58

@joshua_saxe

Joshua Saxe

1 year

Link to the slides for the talk:

Defcon AI Village.pdf

drive.google.com

@joshua_saxe

Joshua Saxe

1 year

Making this deck for my Defcon AI Village keynote took an inordinate amount of time because it meant publicly murdering my darlings: the ~80% of MLsec R&D efforts I worked on over ~10 years and which never reached deployment🧵

Tweet media one

3

13

71

3

20

55

@joshua_saxe

Joshua Saxe

1 year

There's a new class of AI influencers who don't know what a parameter is and are here to teach you about ChatGPT. Maybe this is fine, most AI programmers also don't know how GPUs work.

2

2

56

@joshua_saxe

Joshua Saxe

5 years

What are the temporal dynamics of malware outbreaks? Here's a simple model: an outbreak starts with slow growth, followed by takeoff, followed by a dwindling/taper. Detail image shows Zbot/Zeus, image grid shows other malware families. Work done w/ Giacomo Bergamo.

Tweet media one

Tweet media two

2

14

52

@joshua_saxe

Joshua Saxe

1 year

Giving the keynote at Defcon @aivillage_dc next Friday on what happened in security AI in the decade after AlexNet, and what might happen in the decade after ChatGPT -- Looking forward to seeing many old friends there.

Educating people on the use and abuse of AI

2

10

51

@joshua_saxe

Joshua Saxe

5 years

1\ ML and signature-based detection have been depicted as opposites, when in fact we could just as easily emphasize their similarities. Here's the "condition" block of an ML model generated by my team's ML->Yara compiler. No magic, just a bunch of Boolean logic, like a signature.

Tweet media one

3

21

47

@joshua_saxe

Joshua Saxe

3 months

What's made Yann successful is what's made Elon successful -sticking to first principles against the crowd- in this case open source, open science, and the shocking thesis that LLMs are the wrong architecture to get to human level intelligence

5

5

45

@joshua_saxe

Joshua Saxe

3 months

The rush to collapse generative AI into categories we were already comfortable with -- IQ, interpolation, 'high school level intelligence', n-gram models -- says more about human reflexes and biases than about generative AI's new, alien internal properties

2

4

44

@joshua_saxe

Joshua Saxe

2 years

Why robustness to adversarial examples isn't a first-priority concern on the Sophos AI team.

Tweet media one

8

4

42

@joshua_saxe

Joshua Saxe

3 years

It's a credit to the team that runs MITRE ATT&CK that there seem to be many in infosec who believe that ATT&CK is the main activity MITRE, a sprawling, $2bn/yr quasi-government organization, is engaged in.

3

1

42

@joshua_saxe

Joshua Saxe

2 years

@DavidDeutschOxf You can coax these models into reasoning more clearly. Think of them as sampling an authorial agent from P(agent|text prompt) who is more or less intelligent and rational. Here's GPT3 (text-davinci-003) thinking reasonably about your prompt.

Tweet media one

1

1

41

@joshua_saxe

Joshua Saxe

1 year

There's a huge disconnect between academic adversarial ML work and the battles fought daily by folks operating ML models subject to actual attacks; this paper is such a thoughtful contribution to bridging the gap

Tweet media one

2

13

42

@joshua_saxe

Joshua Saxe

5 years

1/ Some thoughts on the way ML gets talked about in security: Most security problems are not machine learning problems. Like encryption, dual-factor authentication, taint analysis, or hand-crafted IOCs, machine learning is just one of many security tools.

3

8

38

@joshua_saxe

Joshua Saxe

2 years

@EigenGender I don't know, maybe things have changed, a layperson could have come up with "let's think step by step."

2

0

36

@joshua_saxe

Joshua Saxe

5 years

SophosAI YaraML aims to compile production-quality ML detection models to fast & portable Yara rules. Some news: a) @gradientjanitor has just joined the effort! b) We've just released a decent OSX/Mach-O randomforest malware detector @ , 89% TPR/ 0.5% FPR.

Tweet media one

1

16

36

@joshua_saxe

Joshua Saxe

5 years

@EricRWeinstein This feels a bit over the top. The journalism on the Trump White House, the Middle East, tech platforms, and other topics, by NYT, WSJ, and WP, has been excellent. Also, there are exciting new forms like data journalism, Wikipedia, predictive journalism (538), podcasts, etc.

28

1

29

@joshua_saxe

Joshua Saxe

6 months

I don't get why any thoughtful person would be anti-DEI in tech. We are building the technology substrate for the entire world population; scary if we don't actively seek to include everyone

4

7

34

@joshua_saxe

Joshua Saxe

3 months

Super excited to attend CAMLIS and give this talk on the work we're doing to measure and mitigate security risks in models and model/product integrations. Conference should be more interesting than ever with everything going on this year and @ram_ssk 's keynote

@CamlisOrg

Camlis Org

3 months

We are excited to unveil our day two ✌️keynote speaker for #CAMLIS2024 , @joshua_saxe , AI & Security Lead at @Meta . SAVE YOUR SEAT to hear valuable insights from Josh on defense strategies and thwarting the generation of malicious codes by LLMs, & more!

Tweet media one

0

3

9

3

2

35

@joshua_saxe

Joshua Saxe

2 years

@michael_nielsen I use it to make my bad rough text in technical documents better, saying "please rewrite this for brevity and clarity." I also paste in dense text I'm reading and ask it for a TLDR.

0

0

33

@joshua_saxe

Joshua Saxe

9 months

Therapist: "poem poem poem poem poem" Me: "Ok I'm ready to talk about my childhood"

3

2

32

@joshua_saxe

Joshua Saxe

1 year

A theme from discussions at Defcon: "if attacker controlled input goes into an LLM, attacker controlled output comes out of the LLM" needs to become as well internalized in secure app design as "sanitize your SQL queries"

2

7

32

@joshua_saxe

Joshua Saxe

6 years

- My talk at blackhat this year. An intro to deep learning for non-mathy security folks, and a survey of deep learning security work inside and outside of Sophos.

0

9

29

@joshua_saxe

Joshua Saxe

1 year

The idea of intelligence as a one-dimensional number line onto which we can place both machine learning models and living things with no residuals seems to be a major spherical sheep of the x-risk discourse

6

3

30

@joshua_saxe

Joshua Saxe

5 months

The abhorrent revelations about Israel's AI targeting system are making me think about how much of 'AI safety' can't be fixed in technology but only in norms, law, and international agreements around the legitimate uses of AI

3

8

30

@joshua_saxe

Joshua Saxe

2 years

@EigenGender Totally. A 70 year old dream of computer science realized and my non-techie friends are mostly like "oh, cool."

1

0

29

@joshua_saxe

Joshua Saxe

3 years

We arranged our @gather_town avatars in rickroll configuration as part of a send-off to our summer interns. It's this kind of weirdness and creativity that powers innovation on the @SophosAI team :)

Tweet media one

2

5

29

@joshua_saxe

Joshua Saxe

2 years

I've become convinced that in the medium-term, very large models-as-a-service are going to become an important tool within defensive cybersecurity, and this is what Younghoo Lee and I will be talking about at Blackhat USA this year

Tweet media one

1

11

29

@joshua_saxe

Joshua Saxe

7 months

Having been in AI and security for 10+ years, feels like I was in a pond, then a lake, and now I'm in an ocean and it's the calm before the storm

5

5

29

@joshua_saxe

Joshua Saxe

5 years

Large scale malware similarity visualization work by @rpgove , myself, and others. We built a prototype set of analytics and accompanying GUI to accelerate malware analysis over many samples, and did a user study showing efficacy.

Tweet media one

1

11

28

@joshua_saxe

Joshua Saxe

1 month

It's well known in security that untrusted input renders LLM behavior untrustworthy. But a dilemma is that eliminating this risk eliminates much of the value prop of AI, which is why products will keep shipping with prompt injection & adversarial example risks 🧵

1

5

29

@joshua_saxe

Joshua Saxe

5 years

3\ Now let's look at a big malware dataset's compile timestamp behavior. Notice the straight horizontal lines. Those are unique polymorphic hashes reusing the *same* compile timestamp month after month. Also, notice the number of insane back-to-the-future timestamps.

Tweet media one

3

4

28

@joshua_saxe

Joshua Saxe

3 years

We have colleagues on our team in Ukraine, and colleagues who have fled, and there are members of our team housing and helping a Ukrainian colleague and her family. Each barrage of Russian missiles feels deeply personal. I'm heartbroken, and I've never been prouder of @SophosAI .

1

4

28

@joshua_saxe

Joshua Saxe

1 year

Super proud of Meta for open sourcing Llama2. Technologies that work by compressing all of humanity's intellectual labor should be free and open on principle

4

6

29

@joshua_saxe

Joshua Saxe

1 year

Having spent a year in ML outside security, and now being back in security, I'm remembering that perhaps the hardest thing about security ML is identifying actually useful applications of ML. Not so in domains like computer vision where ML is the only game in town

2

3

29

@joshua_saxe

Joshua Saxe

4 years

My team now has a twitter account -- @SophosAI ! If you're interested in tracking #mlsec work from @gradientjanitor , @awalinsopan , @hillarymsanders , myself, @kberlin , @harkervt , @rharang , and others from my team @Sophos , please follow!

1

13

27

@joshua_saxe

Joshua Saxe

3 years

Excited that I'll be giving an RSAC talk in February that will discuss how to assess security vendor ML claims, even if you're not an ML specialist. Looking forward to the talk, and to seeing colleagues and friends in person! The talk will expand significantly on this 👇

@joshua_saxe

Joshua Saxe

4 years

How to evaluate a cybersecurity vendor's ML claims even if you don't know much about ML (thread). 1) Ask them why they didn't solely rely on rules/signatures in their system -- why is ML necessary? If they don't have a clear explanation, deduct a point.

21

304

1K

3

5

28

@joshua_saxe

Joshua Saxe

2 years

@DrCMcMaster @ylecun I just saw some guy explaining what a 'high dimensional space' is to @ylecun and vowed to quit Twitter for a few days :)

3

0

26

@joshua_saxe

Joshua Saxe

5 years

3\ For unknown-unknown attacks, you often want ML anomaly detectors with a human-in-the-loop to follow up on the noisy alerts it'll generate. In sum, I believe you want a complex ecosystem that respects the strengths and weaknesses of both artisanal and mathematical approaches.

2

3

28

@joshua_saxe

Joshua Saxe

5 years

2\ The timestamp field poses a low-key problem for attackers. If they leave the compiler-assigned value they reveal telling details. If they assign a concocted value, their tampering can make them easier to detect. Here's an 'allaple' malware set's random, insane timestamps:

Tweet media one

2

2

25

@joshua_saxe

Joshua Saxe

5 years

1/ Here's a thread on how to build the kind of security artifact "social network" graph popularized by @virustotal and others, but customized, and on your own private security data. Consider the following graph, where the nodes are malware samples:

Tweet media one

1

18

26

@joshua_saxe

Joshua Saxe

3 years

I think MLsec is so confusing mostly because of ML's crazy core promise that it will under certain conditions magically detect bad stuff all by itself, just give it some data! Sounds great, but identifying the conditions in which to use it, and how, is a total mind game.🧵

2

5

25

@joshua_saxe

Joshua Saxe

5 years

@synackpse @IanColdwater No worries! You raised a good point. I should rethink who I follow from an inclusion perspective.

2

0

26

@joshua_saxe

Joshua Saxe

4 years

👇ML Yara rule generated by @SophosAI 's YaraML tool for matching Sunburst / altered Solarwinds PEs; trained using 3 PEs matching FireEye's rule and tested using 3 *other* PEs matching their rule, which it detected successfully ... () ...

Tweet media one

6

13

26

@joshua_saxe

Joshua Saxe

3 years

It's interesting there is such scant academic attention paid to optimizing the fusion of if/then/else logic with ML models, since practically every real-world ML is engaged in said fusion, and lives and dies based on how well it's accomplished.

2

1

26

@joshua_saxe

Joshua Saxe

3 years

I'm refining my idea for a 2nd book: 'Attack Detection Design Patterns'; i.e. how to knit together heterogenous detection tactics into an overall arch + how to operate and tune these systems tactically. Based on real experiences defending actual people. Is this book worth doing?

9

0

24

@joshua_saxe

Joshua Saxe

5 years

What happened to security data visualization? A decade ago, there was lots of interest in it at places like @BlackHatEvents , and good books like @cyberbgone 's and @raffaelmarty 's. But as far as I can tell, visualizations that really exploit preattentive cognition never took hold.

9

5

25

@joshua_saxe

Joshua Saxe

5 years

1\ Let's bypass a convolutional neural network trained to recognize previously unseen bad URLs. The classifier gives a score between 0 (benign) and 1 (definitely malicious). I start by making up a phishing URL: hxxp://wellsfargo-customer-support.webhosting.pl/login

1

6

24

@joshua_saxe

Joshua Saxe

4 months

Why and how ML generalized so much better than anyone expected is one of the core scientific questions of our time, so it's a bit sad when people say "it's just {curve fitting, interpolation, token prediction}" when we've barely scratched the surface

9

4

25

@joshua_saxe

Joshua Saxe

4 years

We've replaced 'blacklist' with 'blocklist' on our website and in our API @ThreatCoalition . Black/whitelist terminology has racial connotations, and also refers to employer practices meant to deprive labor activists of their livelihoods. We can find better terms as an industry.

2

10

24

@joshua_saxe

Joshua Saxe

9 months

Very proud to be part of this team; this is the culmination of a lot of work, and just the beginning of this effort

@AIatMeta

AI at Meta

9 months

Announcing Purple Llama — A new project to help level the playing field for building safe & responsible generative AI experiences. Purple Llama includes permissively licensed tools, evals & models to enable both research & commercial use. More details ➡️

Tweet media one

18

193

878

1

1

25

@joshua_saxe

Joshua Saxe

3 years

Is there a book on detection in security that goes through how to layer rule-based, machine learning, lookup, and reputational approaches optimally? So far it appears these discussions are siloed in the public discourse, and held as folk knowledge inside vendors and large SOCs.

5

0

22

@joshua_saxe

Joshua Saxe

2 years

@GaryMarcus He's perfectly credible as someone trained in ML and computer science. My worry about his trajectory is that he's become essentially a journalist who fawns over and often doesn't ask the hard questions of his (often rich and famous) guests.

1

1

24

@joshua_saxe

Joshua Saxe

3 years

Was feeling listless around work, and then went to the @CamlisOrg conference and feel reenergized, like I paid off some pandemic-era emotional debt and am ready to take on big challenges. I'm newly reminded that face-to-face contact with colleagues is a deep need for me.

2

2

22

@joshua_saxe

Joshua Saxe

2 years

@Carnage4Life I think it only makes sense for reasonable assets like index funds, where the spikes and dips can be seen as variance around a long term historical growth trend.

2

0

22

@joshua_saxe

Joshua Saxe

6 months

'for the benefit of humanity' is an AI industry phrase that hides complexity around who wins, who loses, and how life changes due to this technology

3

5

23

@joshua_saxe

Joshua Saxe

5 years

@tjrwriting @timhwang Here's where I pile on, and reap the fleeting emotional payoff of in-group signaling that comes from dunking on people I've never met

0

1

22

@joshua_saxe

Joshua Saxe

3 years

Just read (& recommend) 'Orange is the New Black' by @Piper . I knew, but didn't really know, about the moral bankruptcy of the US criminal 'justice' system. We're warehousing millions, disproportionately historically disenfranchised, with no actual concern for anyone's betterment

0

2

22

@joshua_saxe

Joshua Saxe

5 years

Dirty laundry in #MLsec : we conflate accuracy in detecting polymorphic variants of known malware, with accuracy in detecting entirely new malware. If we unbundled these problems, we'd reveal that our systems are far less intelligent / resilient to change than has been suggested.

1

3

22

@joshua_saxe

Joshua Saxe

5 years

@bayes_baes @chipro As someone who hires data scientists with ML skills, my experience is that folks who pass my team's hiring bar are actually really hard to find. There are lots of people with these items on their resumes, but far fewer qualified candidates, from what I've seen...

2

0

22

@joshua_saxe

Joshua Saxe

2 months

My mental model is AI valuations probably make sense if AI can plan and reason in the next generation of models, but are a big bubble if not. Exception for media-heavy AI bets like AI marketing / ad creation / music generation, where the value prop is less speculative

2

2

21

@joshua_saxe

Joshua Saxe

6 years

Our paper on malicious HTML detection, to be presented the S&P deep learning workshop: @decodyng @rharang @hillarymsanders . Format agnostic features, novel hierarchical NN architecture for detecting malicious snippets embedded in otherwise banal HTML...

1

12

19

@joshua_saxe

Joshua Saxe

7 years

Thanks! The book's actually by both @hillarymsanders and me. We're very much looking forward to helping more folks in the security community get initiated to data science

@snkhan

Sajid Nawaz Khan

7 years

Really looking forward to “Malware Data Science” by @joshua_saxe . Combines my love of Malware and Data Science!

Tweet media one

0

2

4

4

6

21

@joshua_saxe

Joshua Saxe

6 months

As the set of valid responses to an LLM prompt narrows, the LLM does worse and worse, which is why open problems like conversation and summarization work so well, but anything requiring an exact solution requires endless prompt hacking or just giving up.

4

5

21

@joshua_saxe

Joshua Saxe

7 years

Our paper, "A Deep Learning Approach to Fast, Format-Agnostic Detection of Malicious Web Content", accepted into the Security & Privacy deep learning workshop! ( @decodyng @rharang @hillarymsanders @Sophos ) -- wait for an arXiv version once we're camera-ready ...

0

6

21

@joshua_saxe

Joshua Saxe

5 years

Big news: the Sophos AI team is building a data engineering team in Budapest. Looking for folks interested in building cloud ML infrastructure for research, production training, and operational monitoring, at all levels of seniority.

1

7

21

@joshua_saxe

Joshua Saxe

1 year

Impressive results in this 'Distilling step-by-step' paper where they distill an LLM into small models by predicting both class and step by step reasoning text; "our 770M T5 model outperforms the 540B PaLM model [on one benchmark, due to this procedure]"

3

2

20

@joshua_saxe

Joshua Saxe

5 years

3\ This thread is meant to announce the project and gauge interest -- if there's enough interest we'll take this up as a open source serious effort @ Sophos AI and release the compiler. For now, I'll be releasing a new Yara ML model every week to get feedback and improve.

3

0

20

@joshua_saxe

Joshua Saxe

3 years

If a security ML detection pipeline uses only unsupervised learning I tend to mistrust it. If its creators think it works, they must have had labels to test it; why not use the labels in training at some stage? And if they had no labels, how do we know the system is efficacious?

2

1

18

@joshua_saxe

Joshua Saxe

2 years

@repligate This sounds like the opening of a William Gibson novel :)

1

0

19

@joshua_saxe

Joshua Saxe

5 months

@danwilliamsphil Well put. To me at its best continental philosophy is well observed prose poetry and can be pleasurable / useful as such

1

1

20