TLDR; Upgrade iTerm2
Terminals that implement Kitty graphics need to be extremely careful about what they echo back to the PTY. Here the error response from a non-existent file path leaked unsanitized newlines straight into shell input.
The 3.7.9 patch sanitizes all control characters in Kitty image, drag-and-drop, and title report paths -exactly the right scope.
Other terminals (Ghostty, WezTerm, Kitty itself) don’t appear to share this exact flaw, but it’s a good reminder for anyone building on the graphics protocol.
🚨 A newly disclosed iTerm2 vulnerability allows attackers to execute commands on macOS simply by tricking users into viewing a malicious text file.
Security researcher ii4gsp demonstrated how hidden escape sequences embedded in ordinary files, including README.md, .txt, and .log files, can trigger arbitrary command execution when displayed using commands like cat.
The vulnerability abuses iTerm2's Kitty image protocol, allowing specially crafted terminal responses to inject commands directly into the user's shell. The exploit works with default settings across bash, zsh, and fish, without requiring any additional configuration.
A working proof-of-concept has been published, and the developer has confirmed the vulnerability. All versions supporting the Kitty image protocol before 3.7.9 are affected.
iTerm2 3.7.9, released November 20, includes a security fix. Users should update immediately.
Source: https://t.co/MyZNGejrKS
🚨 A newly disclosed iTerm2 vulnerability allows attackers to execute commands on macOS simply by tricking users into viewing a malicious text file.
Security researcher ii4gsp demonstrated how hidden escape sequences embedded in ordinary files, including README.md, .txt, and .log files, can trigger arbitrary command execution when displayed using commands like cat.
The vulnerability abuses iTerm2's Kitty image protocol, allowing specially crafted terminal responses to inject commands directly into the user's shell. The exploit works with default settings across bash, zsh, and fish, without requiring any additional configuration.
A working proof-of-concept has been published, and the developer has confirmed the vulnerability. All versions supporting the Kitty image protocol before 3.7.9 are affected.
iTerm2 3.7.9, released November 20, includes a security fix. Users should update immediately.
Source: https://t.co/MyZNGejrKS
I don’t think “multiplexer” is the right term.
The original idea was to use a single TTY to multiplex multiple terminal sessions. I know @superlogical brought new life to the term, but what we’re actually building are terminal window managers.
Musk is wrong here, of course. What he says is untrue and Kremlin propaganda. I was on Maidan, every day from start to finish. (I even wrote a book about it.) There are few people who can organize and mobilize like Ukrainians. The CIA wishes it could foment uprisings as quickly and efficiently as Ukrainians can.
The simple facts are:
Ukrainians came out to protest Yanukovvych's turn away from the EU and towards Russia.
Those protests turned to large-scale uprising after violence perpetrated by Yanukovych's security forces.
It swelled into a revolution by Ukrainians who were fed up with state violence and worried about the direction Yanukovych was taking their country.
Yanukovych agreed a deal for power sharing and early elections, but then fled.
After Yanukovych fled and abandoned his constitutional duties, Ukraine's parliament voted to remove him.
Moreover, Putin and Russia began preparing for the invasion of Ukraine before Yanukovych was ousted. You can look at the Kharkiv Pact in 2010, and the start of operations in February 2014, before Yanukovych fled.
Something I learned a bit ago that I think needs telling again...
LLMs are far more deterministic than most people realize.
During inference, a pseudorandom seed drives token sampling, creating variability in responses.
Lock the seed and the model reproduces the exact same response, token for token.
Same inputs. Same math. Same output.
What looks like independent thought is often just pseudorandom sampling.
Hi I'm Breck and I once saved entire internet from going down.
In 2014 Microsoft and NudgePad merged and I joined Azure team, Microsoft's cloud.
One time I was assigned responsibility for making Azure more reliable. At time we had one 9 reliability, which, although record for a Microsoft product, was not good compared to Digital Ocean's 99.99% reliability.
So I started by building a near realtime dashboard to see all Azure errors, then I drilled in on each error, one at a time, with most frequent first.
After about a week or two of tracing down most common errors and getting those fixed, I started scanning for less frequent but nonetheless interesting errors.
One in particular caught my eye, because it had a lot of chinese characters in error message.
I translated the chinese (I had taken 1 semester of Chinese my freshman year at Duke, and also I used Google Translate), but it was just some generic domain name.
So I then tried to recreate the error by setting up my own subdomain with that Chinese combo, and lo and behold I was able to reproduce.
Then I started digging, and found deep in some Azure code, a Javascript "eval", that ran in the logged in Azure user's browser and was fed in part by data passed into query string in a simple GET request.
My arm hairs are starting to go up again as I sit here and rememeber how the implications slowly dawned on me.
IIRC I then pulled in my colleagues Mo, Alvaro, Chris, and my boss David and we dug in further.
I think everyone realized this was serious, but I don't think even they grasped the severity.
I then quietly did a test where I put an <img tag on a website with a GET request to azure containing a malicious query string, and then I visited that web page from a logged in Azure account and watched as it deleted all my VMs.
I knew at that moment I was not going to be making my normal 2pm swim at pro club, 3pm ping pong game with Joe and Jacob and Aaron, and 4pm flag football game with Taylor Swift Fan Club.
We quickly worked on a patch, and before it was too dark had it running through CI and headed for prod.
In the meantime I did another experiment, where this time the malicious URL didn't nuke someone's account, but instead hijacked their account to embed these malicious <img tags on all their azure websites.
This thing could have gone viral, and not in the fun Hawk Tuah way but more like in the half the internet dies way.
Anyway, around 3am or so I verified fix was in, that there was no malicious uses (AFAICT), and my boss approved a cab ride home.
It wasn't always a country club in Redmond!
Sorry for the novel. Thank you for coming to my Ted Talk.
A hui Hou.
Breck
We're back to Amsterdam for Greeks in Tech on Wednesday, December 9! 🇳🇱🇬🇷🌷
Excited to return to the Netherlands in our quest to meet with community peers around the world.
Spread the word, tell your friends, and see you there!
Option 1: He's not the smartest person in the world, that's fine (who is, anyway), just don't take what he says as godsend next time.
Option 2: He'll just move where power shifts to promote his business interests. Many businessmen do, just don't take what he says at face value.
Introducing SRE-1 a 27B model that dominates the Pareto frontier of accuracy/cost on production debugging, beating frontier models in a coding-adjacent domain
Built by 11 people on open weights models + domain expertise. Final run: 7.4 days on 512 GPUs https://t.co/9IzysHZEmL