sagitz Profile
sagitz

@sagitz_

Followers
4,307
Following
736
Media
46
Statuses
199

Cloud Security Researcher at @wiz_io • Microsoft Most Valuable Researcher 21/22/23 • Black Hat Speaker

Joined March 2019
Don't wanna be here? Send us removal request.
Explore trending content on Musk Viewer
Pinned Tweet
@sagitz_
sagitz
2 months
We found a Remote Code Execution (RCE) vulnerability in @Ollama - one of the most popular AI inference projects on GitHub. Here is everything you need to know about #Probllama (CVE-2024-37032) 🧵👇
Tweet media one
23
355
2K
@sagitz_
sagitz
1 year
We found two 0-day vulnerabilities in @Ubuntu kernel and it all started by reading descriptions of old CVEs 📖 Thread about the discovery of #GameOverlay 🧵👇🏼
Tweet media one
17
462
2K
@sagitz_
sagitz
5 months
We uploaded a backdoored AI model to @HuggingFace which we could use to potentially access other customers’ data✨ Here is how we did it - and collaborated with Hugging Face to fix it 🧵⬇️
Tweet media one
17
249
2K
@sagitz_
sagitz
2 years
Hex Rays + OpenAI wen?
Tweet media one
Tweet media two
Tweet media three
Tweet media four
32
237
1K
@sagitz_
sagitz
3 years
I'm excited to share information about our research, in which we (+ @nirohfeld ) found a critical vulnerability in Azure Cosmos DB itself - effectively allowing malicious actors to fully compromise databases of thousands of customers. @wiz_io #ChaosDB
8
122
234
@sagitz_
sagitz
3 years
Recently @nirohfeld and I started looking at #Azure . Perhaps the most severe vulnerability I have ever reported. Stay tuned🤩 #MSRC #BugBounty
Tweet media one
13
37
210
@sagitz_
sagitz
3 months
We discovered that by uploading a malicious AI model to @Replicate , a leading AI-as-a-Service platform, we could read and modify prompts of other customers 🤯 Here is exactly how we did it 🧵⬇️
Tweet media one
7
63
211
@sagitz_
sagitz
2 months
The issue is a simple Path Traversal vulnerability which can be exploited by pulling a model from a private registry. By specifying a malicious digest field, it is possible to overwrite any file on the system!
Tweet media one
3
18
166
@sagitz_
sagitz
2 years
Found it faster than Jackalope or AFL👀
Tweet media one
Tweet media two
6
24
160
@sagitz_
sagitz
5 months
Kudos to @HuggingFace security and infrastructure team who fixed these issues in record time, and implemented additional security measures to prevent this from happening in the future🤗🙌
1
0
117
@sagitz_
sagitz
5 months
Interested in the technical details of this research? Check out our blog in which we detail the issues we found and how they can be mitigated 🔗⬇️
1
6
101
@sagitz_
sagitz
2 months
This behavior is a common pattern we see in AI tooling: an immature codebase with simple vulnerabilities, no common security mechanism out-of-the-box. Infrastructure security is one of the most challenging aspects of AI security.
1
6
102
@sagitz_
sagitz
5 months
After establishing a foothold inside HF’s infrastructure, we quickly noticed that we were running inside a Kubernetes pod hosted on AWS. A couple of EKS tricks later, we were able to escalate our privileges in the cluster and potentially take over the service💥
Tweet media one
2
4
99
@sagitz_
sagitz
11 months
I typically try to exclusively discuss security research on Twitter and especially refrain from discussing politics. But I will not stay indifferent to this situation. I live in Israel, and in the last week, my country has been forced into a blood-soaked war. On October 7th,
1
21
99
@sagitz_
sagitz
2 months
We exploited the vulnerability by overwriting /etc/ld.so.preload to load our malicious shared library. This escalated our Arbitrary File Write into a Remote Code Execution 😎
5
0
88
@sagitz_
sagitz
2 months
It's a bad idea to expose your Ollama instance anyway. Attackers can leak models, modify prompts, and use compute resources even without exploiting a vulnerability. It literally takes one cURL command to inject a malicious prompt into existing Ollama models. It's a feature.
Tweet media one
2
5
85
@sagitz_
sagitz
2 months
Although a patch for this issue (0.1.34) has been available for over a month, most publicly exposed instances found on Censys are still vulnerable🤯
1
1
79
@sagitz_
sagitz
1 year
Both of these vulnerabilities were very easy to exploit, in fact, our exploit code is mostly bash 🙈 We wrote our PoCs and reported everything to the Ubuntu team, who were quick to fix the issues and assigned themCVE-2023-2640 and CVE-2023-32629
Tweet media one
1
5
74
@sagitz_
sagitz
5 months
AI Models can come in different formats, based on the framework they were developed in. Some formats are safe, while others (like Pickle) allow Remote Code Execution as a feature!
Tweet media one
2
2
64
@sagitz_
sagitz
5 months
Hugging Face, one of the best-known AI-as-a-Service providers, conveniently lets users interact with the AI models hosted on their platform using their own inference infrastructure. This feature is called Inference API.
Tweet media one
1
0
60
@sagitz_
sagitz
4 years
What's written: "PHP 8 introduces two JIT compilation engines" What I read: "PHP 8 introduces new attack surface"
1
8
53
@sagitz_
sagitz
1 year
Our journey started when our team at @wiz_io read the advisory about CVE-2023-0386, a local privilege escalation in the Linux kernel. The vulnerability exploited OverlayFS to copy SUID files from a nosuid mount to outside directories, enabling privilege escalation to root.
Tweet media one
2
2
51
@sagitz_
sagitz
2 years
OpenAI corpus generation™
Tweet media one
1
3
50
@sagitz_
sagitz
5 months
We were wondering: What would happen if we uploaded a malicious (pickle) model to Hugging Face and interacted with it using Inference API? Would our code be executed? Would our model share the same infrastructure as other Hugging Face users? 🤔
1
0
46
@sagitz_
sagitz
5 months
We took a legitimate model (gpt2) and modified it to execute shell commands when encountering the word ‘Backdoor’ in the prompt.
Tweet media one
Tweet media two
1
0
43
@sagitz_
sagitz
3 years
Barely made it to MSRC 2021 Q3 Security Researcher Leaderboard😅 Congrats to all other researchers on the list! #bugbounty #msrc
Tweet media one
6
2
41
@sagitz_
sagitz
1 year
What’s the difference? Each of us was running on a different kernel version. Apparently, Ubuntu made changes to OverlayFS a while back. In certain kernel versions, file capabilities are copied as-is, and in some, they are correctly converted relative to the current user namespace
Tweet media one
1
2
42
@sagitz_
sagitz
1 year
After finding the first 0-day, we tried the opposite approach - where else are file capabilities copied without conversion? 🔍 We found another function that is accessible via a slightly different code path. We modified our exploit code and voila - 0-day #2 ! 🔥
1
0
41
@sagitz_
sagitz
1 year
To bypass the patch, we tried to reproduce the same vulnerability, but using file capabilities instead of SUIDs. IT WORKED!🤯 Well, on my machine. But when we tried the same exploit on @shirtamari 's machine, it failed. Weird🤔
1
0
40
@sagitz_
sagitz
1 year
This should perfectly mitigate the SUID exploit, as SUIDs must be owned by root to be effective. However, SUIDs are not the only way to elevate privileges. There's also file capabilities - a way to grant root-like capabilities to a file without needing it to be owned by root.
Tweet media one
1
1
38
@sagitz_
sagitz
1 year
To mitigate the issue, an additional check was added to verify that the owner of the modified file is present in the current user namespace:
Tweet media one
1
0
36
@sagitz_
sagitz
2 years
I will be in Vegas for Black Hat and Defcon, if you see me, come say hi! Changed my profile pic so you'll know what I look like :)
Tweet media one
0
0
34
@sagitz_
sagitz
3 years
Google knows me too well
Tweet media one
0
0
31
@sagitz_
sagitz
2 years
Excited to be featured on MSRC 2022 Q1 Security Researcher Leaderboard! Got #15 overall and #5 in Azure 🤩 Thank you @msftsecresponse and congratulations to everyone else on the list :)
Tweet media one
@msftsecresponse
Security Response
2 years
Congratulations to all the researchers recognized in this quarter’s MSRC 2022 Q1 Security Researcher Leaderboard! For more information, check out our blog post: #cybersecurity #bluehat #msrc
0
9
45
1
1
30
@sagitz_
sagitz
3 years
If you are at @BlackHatEvents tomorrow, come check out our (+ @nirohfeld ) talk: "ChaosDB: How We Hacked Databases of Thousands of Azure Customers." 11:20am-12:00pm (Room BC, ICC Capital Suite 12, Level 3)
Tweet media one
1
1
28
@sagitz_
sagitz
1 year
Who knew that Bing is such an attractive target for research? 🧐 AAD Misconfiguration ➡️ XSS in Bing ➡️ O365 Tokens 🤯
@hillai
Hillai Ben-Sasson
1 year
I hacked into a @Bing CMS that allowed me to alter search results and take over millions of @Office365 accounts. How did I do it? Well, it all started with a simple click in @Azure … 👀 This is the story of #BingBang 🧵⬇️
Tweet media one
267
3K
16K
0
1
27
@sagitz_
sagitz
3 years
We (+ @nirohfeld ) will be presenting our research about #ChaosDB at #BlackHatEU - See you there!🤩 @BlackHatEvents
Tweet media one
0
2
27
@sagitz_
sagitz
1 year
Beautiful research and amazing vulnerabilities🔥 If you've ever wondered what a real life adversarial attack on Kubernetes looks like - this is a must read💯
@ronenshh
Ronen
1 year
Today we share our Alibaba Cloud research for the first time, where we gained unauthorized access to other customers' databases in two different services 🚨 This complex research involved RCE, PE, Container escape, K8s lateral movement, and supply chain attack. Check it out 🧵
Tweet media one
13
415
1K
0
3
22
@sagitz_
sagitz
2 years
The exact moment I realized I had become dependent
Tweet media one
4
0
19
@sagitz_
sagitz
11 months
It's scary how easy it is to make this mistake and accidentally expose a storage account, even in 2023 🫣
@hillai
Hillai Ben-Sasson
11 months
We found a public AI repo on GitHub, exposing over 38TB of private files – including personal computer backups of @Microsoft employees 👨‍💻 How did it happen? 👀 A single misconfigured token in @Azure Storage is all it takes 🧵⬇️
Tweet media one
59
633
3K
1
0
19
@sagitz_
sagitz
1 year
Very interesting campaign that is still affecting many well-known vendors. Two things I can’t figure out: 1. where did the attackers obtain the credentials from? 2. with such a powerful capability, why did they choose to redirect to adult sites instead of keeping it covert?
@AmitaiCo
Amitai Cohen 🎗️
1 year
Yesterday we published a report on a curious website hijacking campaign we've been monitoring since last October. Our investigation turned up ~10K hacked websites, a few loose ends, and IOCs and samples to check for signs of infection and research further.
0
10
25
0
1
18
@sagitz_
sagitz
2 years
A detailed blog post regarding the supply-chain vulnerability we discovered in IBM Cloud Databases for PostgreSQL. Don’t miss it!🔥 #cloud #security #HellsKeychain
@ronenshh
Ronen
2 years
Excited to share #HellsKeychain , a supply-chain attack affecting IBM Cloud Databases for PostgreSQL. Read about how we chained multiple vulnerabilities in IBM Cloud’s environment to intervene in their internal image-building process
3
10
59
0
2
17
@sagitz_
sagitz
2 years
Incredible to see how quickly Azure adapted their threat model to address cloud vulnerabilities, particularly cross-tenant vulnerabilities. Our research is making a real impact! 😊
@markrussinovich
Mark Russinovich
2 years
The first of a blog series on how we approach Azure cloud service security: "Microsoft Azure's defense in depth approach to cloud vulnerabilities"
3
79
228
1
2
17
@sagitz_
sagitz
1 year
This CTF is a great intro to IAM misconfigurations in particular and cloud security in general. Challenge #6 is my favorite🤭 Try it yourself ➡️
@nirohfeld
Nir Ohfeld
1 year
Think you are an AWS IAM expert? 🤖 Put on your attacker hat and play our new CTF: The Big IAM Challenge! 🎉
1
20
88
0
1
16
@sagitz_
sagitz
2 years
This is the regex that we found in pg_ident.conf that our CN must match to successfully authenticate. Can you spot the bug? The regex ends with a wildcard! will match regex and therefore will let us authenticate! (9/n)
Tweet media one
1
6
16
@sagitz_
sagitz
3 months
We will be talking about this issue (and other AI vulnerabilities we discovered) in our upcoming @BlackHatEvents session - make sure to check it out!
Tweet media one
1
0
16
@sagitz_
sagitz
3 years
@nirohfeld @wiz_io The email we received from Microsoft
Tweet media one
1
2
15
@sagitz_
sagitz
2 years
At DEFCON playing an infosec card game. Hit me up if you want to play :)
Tweet media one
0
1
14
@sagitz_
sagitz
3 years
I have seen the exploit and it is beautiful 😍
@nirohfeld
Nir Ohfeld
3 years
Love working with @msftsecresponse . Another critical Azure vulnerability. This time an RCE 🤩 #BugBounty #Azure @wiz_io
Tweet media one
14
46
318
0
0
13
@sagitz_
sagitz
3 months
Interested in the technical details of this research? Check out our blog in which we detail the issues we found and how they can be mitigated 🔗 ⬇️
1
1
12
@sagitz_
sagitz
3 months
We used rshijack to inject a raw TCP packet into the existing (already-authenticated) TCP connection. It worked!😎 We used this to gather information about the Redis instance and we learned that it stores the prompts and results of other Replicate's customers! 🤯
Tweet media one
2
0
10
@sagitz_
sagitz
2 years
Absolutely brilliant blog post by @chompie1337
@chompie1337
chompie
2 years
Thrilled to share my new blog post: Put an io_uring on it: Exploiting the Linux kernel. Follow me while I learn a new kernel subsystem + its attack surface, find an 0day, build an exploit, + come up with some new tricks. I go deep and demystify the process
47
637
2K
1
0
11
@sagitz_
sagitz
2 years
Hexacon fomo is real 🫠
1
0
11
@sagitz_
sagitz
2 years
And an SQLi. The future looks bright :)
@ProgrammerDude
Arian van Putten
2 years
Copilot writing an authentication page in PHP using MD5. Great
Tweet media one
142
224
2K
0
0
11
@sagitz_
sagitz
3 months
We wrapped everything up and reported the issue to Replicate. It was immediately clear how seriously they take security matters. We scheduled a call to discuss possible mitigations for the issue and the full mitigation was deployed the next day 👑
1
0
10
@sagitz_
sagitz
3 months
After playing around with Replicate's AI model format (called Cog), we managed to craft a model that can execute arbitrary shell commands on request. We uploaded it to Replicate and got Remote Code Execution!💥
Tweet media one
2
0
10
@sagitz_
sagitz
2 years
At Black Hat Europe 2021, we presented our research of #ChaosDB - a major cross-account issue in Azure Cosmos DB. Here’s the research as a refresher () Following Black Hat, we wondered – could we reproduce a similar issue in other Azure services? (2/n)
1
1
9
@sagitz_
sagitz
3 months
Finally, we injected a Lua script into the centralized Redis that modified the inputs and outputs of another one of our accounts, which proved that we could interfere with other customers' prompts and predictions!
Tweet media one
1
0
8
@sagitz_
sagitz
1 year
What if I told you that you could be productive by watching YouTube at work? Can't wait for the next episode!
@wiz_io
Wiz
1 year
🎧 New "Crying Out Cloud" ep! Discover cloud security news with Eden & @AmitaiCo 🌩️ In this episode: Mysterious redirections, fake ads, gaming breaches & DoD leaks 🎮🔓 Watch now: 🍿📺 #cloudsecurity #podcast
0
2
14
1
0
9
@sagitz_
sagitz
7 months
Does anyone have any contact to share with the security folks at @replicate ? 🙏
1
2
9
@sagitz_
sagitz
2 years
This thread is jam-packed with valuable information🔥
@nirohfeld
Nir Ohfeld
2 years
💡 Two key takeaways from this thread: 1. When attacking K8s, target imagePullSecrets for access to container registry credentials. 2. Always scan container images (including layers and metadata files) for secrets. #cybersecurity #k8s #containers #securitytips
1
7
33
0
2
9
@sagitz_
sagitz
2 years
Microsoft fixed this issue in three days. Thanks to Microsoft’s outstanding collaboration with researchers, vulnerabilities like these are responsibly disclosed and fixed instead of being exploited by malicious actors. They even awarded us a $40K #BugBounty (11/n)
Tweet media one
1
1
9
@sagitz_
sagitz
9 months
This blows my mind. @xenovacom 's work on Transformers.js is truly revolutionary. Utilize AI models directly in your browser, on a commodity CPU, without the requirement of a GPU 🤯
@xenovacom
Xenova
9 months
Distil-Whisper small is finally here! 🔥 Over 10x smaller, 5x faster, and within 3% WER of large-v2. 🤯 Since it's only 166M params, it can even run locally in your browser with 🤗 Transformers.js! Check it out! 👉
9
62
354
0
1
8
@sagitz_
sagitz
3 years
quality content
@gamozolabs
Brandon Falk
3 years
My keynote from SSTIC is now up on YouTube! Enjoy, had a lot of fun animating and editing this!
7
33
145
0
0
8
@sagitz_
sagitz
3 months
We investigated the network and noticed an already-established TCP connection handled by a process in a sidecar container. We used tcpdump to examine the contents of this TCP connection, and realized it was a plaintext Redis protocol
Tweet media one
1
0
7
@sagitz_
sagitz
3 years
Another great find by @nirohfeld . Beautiful and easy-to-exploit vulnerability with CVSS 9.8/10. "...strange 'Enhanced Security' commit was introduced on August 12th 2021 ... Microsoft’s official patch (v1.6.8-1) was only released on September 8th 2021" 🤔
@nirohfeld
Nir Ohfeld
3 years
Microsoft just patched 4 vulnerabilities we ( @wiz_io ) recently reported, including a CVSS 9.8 RCE. These vulnerabilities affect countless machines as the OMI agent is silently installed when enabling many Azure services. #PatchTuesday
5
129
276
0
0
8
@sagitz_
sagitz
2 years
Luckily for us, Azure PostgreSQL Flexible Server does not run stock PostgreSQL. After researching the modifications Azure introduced to the engine, we found a way to elevate our privileges to superuser. This was sufficient for us to invoke a reverse-shell! (5/n)
Tweet media one
2
0
7
@sagitz_
sagitz
1 year
@OphirHarpaz :q! ↑ ctrl+a ⌦ ⌦ ⌦ nano ↵
0
0
7
@sagitz_
sagitz
2 years
Chaining these two vulnerabilities together effectively allowed us to gain read access to the PostgreSQL (Flexible) databases of other Azure customers. Just imagine a scenario where a malicious actor exploits this issue to access the databases of Fortune 500 companies. (10/n)
1
0
6
@sagitz_
sagitz
3 months
Having RCE as root within our own container on Replicate's infra, we noticed that we were running inside a K8s cluster hosted on GCP. But our pod was not privileged and we did not have any service account attached, limiting our possibilities for escaping the container 🥲
1
0
5
@sagitz_
sagitz
2 years
Now we were running in a dockerized environment with a recent kernel. No easy docker escape this time. But we did have a couple of interesting network interfaces. More interestingly, we could connect to other hosts in our subnet on port 5432 (PostgreSQL port). (6/n)
Tweet media one
1
1
5
@sagitz_
sagitz
3 years
PHP❤️
Tweet media one
0
2
5
@sagitz_
sagitz
2 years
@steventseeley Great initiative! 🔥 Recent: Unauthenticated Database Takeover + RCE in Azure Cosmos DB - All time proud: Windows DNS Server RCE (Memory Corruption) - Excited to read everything on this thread🤩
0
0
5
@sagitz_
sagitz
2 years
Our criteria for the research target were the following: 1. Service must be managed and provide customers with a dedicated instance 2. We need to execute arbitrary code, either as a feature or through vulnerability 3. Popular service that contains sensitive data (3/n)
1
0
5
@sagitz_
sagitz
3 months
We tried connecting to this Redis instance, but it required authentication 🧐. However, we already had an authenticated, plaintext, active session within our network namespace. And we had all the network capabilities! What if we inject raw packets into that TCP stream? 👀
1
0
5
@sagitz_
sagitz
2 years
We continued our recon and examined the PostgreSQL configuration files. We found that the replication user of the database is allowed to connect from 10.0.0.0/8. Maybe we can connect to other customers using the replication user? (7/n)
Tweet media one
1
0
5
@sagitz_
sagitz
4 years
@chompie1337 Very nice! 🤩
1
0
4
@sagitz_
sagitz
3 years
@chompie1337 Wow, amazing writeup🔥
1
0
4
@sagitz_
sagitz
2 years
@_L4ys You can get specific versions of binaries (for patch-diffing) from
0
0
3
@sagitz_
sagitz
4 years
@sleepya_ @33y0re Incredible job!👏
0
0
3
@sagitz_
sagitz
2 years
@0xdea Another example :)
@sagitz_
sagitz
4 years
CVE-2020-1350: A cool Windows DNS Server vulnerability (2003->2019) we found at @_CPResearch_ and got patched today #PatchTuesday
5
55
128
0
1
3
@sagitz_
sagitz
1 year
@TalBeerySec Found a remote code execution issue, decided not to report because the exploitation scenario is so unlikely it would ultimately be just a waste of dev time.
0
0
3
@sagitz_
sagitz
2 years
What are the username and password for the replication user? Turns out that there are none. To use the replication user, you must authenticate with Client Certificate Authentication. (8/n)
1
0
3
@sagitz_
sagitz
2 years
Azure PostgreSQL Flexible Server fits our criteria perfectly. It is popular, contains sensitive data and offers functionality to execute OS-level commands. But when we attempted to run shell commands, we lacked privileges. (4/n)
Tweet media one
1
0
3
@sagitz_
sagitz
8 months
@replicate can you check your DM? 🙏
0
0
2
@sagitz_
sagitz
3 years
1
0
2
@sagitz_
sagitz
3 years
@EyalItkin @_CPResearch_ @ynvb It was a pleasure working with you! Good luck on your new adventure :)
1
0
2
@sagitz_
sagitz
2 years
Huge props for that! This is a significant step towards better security for Office Next step: bounties :)
@secbughunter
Tom Gallagher
2 years
Hope everyone enjoys tomorrow's Office symbol release. Thanks to everyone inside of Microsoft for making this happen. It was a bit of work. We look forward to enabling high quality Office research. 😀
9
76
296
0
0
2