Uri Tzadikevitch retweeted
Coding agents write a change in minutes. Testing it on AWS ECS still takes hours.
Every team on ECS we've talked to describes the same loop: build the image, push it, register a new task definition revision, wait for the rollout, check what happened, and do it all again if it's wrong.
That was already too slow for humans, but it's so much worse for agents who write a change in seconds, then have to wait a full rollout to learn whether it worked.
Most of the speed you got from the agent ends up spent on waiting for rollouts.
On Kubernetes, mirrord fixed this a while ago. 20,000+ developers use it along with their AI agents to test generated code against their real staging environment without deploying it.
Today we're going beyond Kubernetes with the launch of mirrord for ECS.
• Your code runs on your laptop or in an agent's sandbox, but it gets real traffic, environment variables, network access, and files from your ECS task in staging. You see if a change works in seconds, not after a rollout.
• To set it up, it's just one task definition change.
• Only requests that carry your session header come to your code, so you, your teammates and your agents can each test your own version of the service at the same time, on the staging environment you already share.
mirrord for ECS is live today, and it’s free to get started with! Let your agents test against real staging without deploying, and without getting in anyone else's way.
Uri Tzadikevitch retweeted
🚨 Cloudflare Containers flaw let customers read other tenants' leftover databases and credential files.
Researchers recovered residual data on 18 of 24 production placements by reading reused disk blocks. Cloudflare fixed the flaw and says it found no evidence of abuse beyond authorized testing.
How it worked → thehackernews.com/2026/09/cl…
Uri Tzadikevitch retweeted
The craziest sandbox escape we've found so far:
Our team at @Accomplish_ai exploited a vulnerability in @Cloudflare Containers that let a sandbox read files left behind by other customers' containers. Now fixed.
Full write-up on the @CloudflareDev blog: blog.cloudflare.com/containe…
Uri Tzadikevitch retweeted
What's worse than escaping a sandbox and reading the host's files?
Escaping Cloudflare's, and reading other customers' files.
How did that happen? Cloudflare Containers' disk is thin-provisioned, which means the host keeps a shared pool of storage for every container on the machine, and attaches it lazily in 64 KB chunks, instead of provisioning it all upfront.
When we tried reading a chunk of unallocated disk space in a container, we got zeros, which is to be expected.
But we noticed something strange: the chunks that we already wrote to were full of bytes we didn't write.
Turns out a write is what makes the thin provisioner allocate the chunk, and Cloudflare wasn't zeroing it on the way in, so the rest of the chunk still held leftover data from the previous customers' containers that happened to run the same machine.
By writing 4 KB to every chunk of the disk. The other 60 KB of each chunk were instantly filled with other customers' files: .env, credential files, .pem keys, SQLite databases, Chromium cookies and localStorage.
I've disclosed vulnerabilities to many vendors over the years, but working with Cloudflare was something else. Putting aside the speedy triage, mitigation, and generous bounty, they have a unique and earnest engineering culture of transparency.
Which lead to us collaborating and writing the full postmortem together:
blog.cloudflare.com/containe…
Uri Tzadikevitch retweeted
Very nice find from the Accomplish security team. They found a way to read other customers' data on Cloudflare Containers.
When one container was deleted, its disk space was handed to the next container without being wiped first. So a new container could sometimes read what the previous one had left behind, including full databases.
Very cool work!
blog.cloudflare.com/containe…
Uri Tzadikevitch retweeted
Our research team found a disk isolation bug in Cloudflare Containers, the platform underneath Cloudflare Sandboxes, that exposed other customers' files from inside a sandbox, including credential files and browser cookie stores.
Reported Sept 4, fixed fast, disclosed jointly with @Cloudflare today.
Security researchers at Accomplish identified a Cloudflare Containers vulnerability involving residual disk data. We've detailed the issue, investigation, and fix in our latest post. The vulnerability is fully remediated, and no customer data was compromised. cfl.re/4rqPpiW
Uri Tzadikevitch retweeted
Security researchers at Accomplish identified a Cloudflare Containers vulnerability involving residual disk data. We've detailed the issue, investigation, and fix in our latest post. The vulnerability is fully remediated, and no customer data was compromised. cfl.re/4rqPpiW
Our craziest escape yet:
The @Accomplish_ai research team was able to exploit a vulnerability in Cloudflare Containers that let a sandbox read other customers' files - SQLite DBs, Chromium profiles, .env files etc,
Cloudflare Sandboxes and Browser Run run on the same disk implementation and were affected too.
We reported this to @Cloudflare, who super quickly fixed it.
Read @CloudflareDev post in collaboration with Accomplish researcher @orenyomtov on their official blog:
blog.cloudflare.com/containe…
Uri Tzadikevitch retweeted
We escaped Docker's hypervisor with three lines of bash. CVE-2026-77179: A container gets complete read and write access to the host filesystem.
When you mount a folder into a container, Docker's VMM uses virtio-fs, and the file server runs on the host.
Because of a TOCTOU bug, if a container opens a file, deletes it while holding its file handle open, and replaces the parent folder with a symlink, the kernel will follow the symlink to anywhere on the host.
Full technical breakdown:
accomplish.ai/blog/escaping-…
Readers added context they thought people might want to know
This affects Docker Sandboxes on macOS and Docker Desktop only if the new Docker VMM (beta, not default) is enabled; it is a virtio-fs shared workspace symlink/TOCTOU issue, not a general container or hypervisor escape.
docs.docker.com/security/secur…
cve.org/CVERecord?id=C…
docs.docker.com/ai/sandboxes/r…
Shabbat Shalom.
@Docker has issued a critical level (!!) CVE-2026-77179, reported by @Accomplish_ai researcher @orenyomtov.
If you use Docker Sandboxes, run sbx --version. Make sure you’re running 0.42.0 or later.
If you use Docker Desktop, you want 4.88.0 or later
Docker Desktop and Docker Sandboxes are both affected. Docker Desktop is only affected if Docker VMM is turned on in Settings.
It’s a good thing we found it now, because Docker VMM is scheduled to become the default for Docker Desktop at the end of October 2026!
The escape is in Docker's hypervisor for Mac: a container gets complete read and write access to the host filesystem.
Read the full details of this escape in our blog:
accomplish.ai/blog/escaping-…
Uri Tzadikevitch retweeted
We escaped OpenAI's Codex sandbox by dumping V8's shared JavaScript heap and scanning it for the privileged UUID-shaped token belonging to the adjacent trusted V8 context.
I sometimes see cybersecurity as a never-ending cycle of building trust boundaries and finding ways to break them. This is the story of one interesting break.
When OpenAI built this default-enabled MCP server, they ran a trusted, privileged JS context alongside an untrusted one, in the same Node.js process.
They assumed V8's VM contexts could serve as a security boundary, because they do create isolated execution environments, each with its own distinct global object and scope.
But they forgot that these contexts share a single memory heap. So our untrusted code snapshotted that heap, pulled the token out of it, and used it to reach the privileged side.
Read the full technical breakdown at:
accomplish.ai/blog/escaping-…
This is our third sandbox escape writeup, and after Claude Code and Cursor, it's OpenAI's Codex this time!
This one is especially cool, for two reasons:
1. One of the escapes is from the open source codex cli. As an open source project, we found it to be generally much more secure than its competition.
2. The second escape is in the rust-based closed source - and it leverages a sophisticated heap attack.
Kudos to @OpenAIDevs @OpenAI for fixing this super quickly and to @Accomplish_ai researcher @orenyomtov for finding it.
Both fixed now, and here are all of the details:
🚨 SANDBOX ESCAPE ALERT 🚨
SPACEXAI: PWNED 🫡
CURSOR: LIBERATED 🦋
Following a fix shipped by the Cursor team (in about a week, kudos), we present a variant of the Beltdown vuln we shared a few days ago -
Escaping Cursor CLI's sandbox to run arbitrary code on your Mac:
🚨 SANDBOX ESCAPE ALERT 🚨
ANTHROPIC: PWNED 🫡
CLAUDE CODE: LIBERATED 🦋
@Accomplish_ai researchers uncovered an massive RCE/ACE vulnerability that we are publishing today, where an untrusted git repo opened in Claude Code can escape the macOS sandbox and execute commands on your computer as your privileged user, also bypassing the permission prompt >>
And... all hacked 🦇
Thanks @alexrkonrad for covering some of our work: teapawt.net/alexrkonrad/status/209…
Following fixes shipped from all labs, we will be posting a series of more technical posts, depicting the details of the sandbox escapes, starting with 🥁
Beltdown: an RCE made possible due to a Claude Code Sandbox escape:
teapawt.net/_orcaman/status/209816…
Uri Tzadikevitch retweeted
i just won $200k using this one prompt
a new critical vuln was found by AI augmented researchers, patched by us 1hr later
this time it was quite twisted, took me a good amount of time to grasp
it was never used to attack us, as far as we can dig
I'm now convinced that even well designed/written software can have so many such small, twisted issues
it takes good experience with vulnerabilities and long chains of actions to suspect, find and exploit them
10yoe+ and solid CS knowledge is not enough
beyond the classic security footguns, that we, software engineers, know about and try hard to prevent
some class of issues, often more case-by-case, can be just invisible to us, builders, but visible to security focused folks
don't be too confident, open a bounty, let yourself be humbled, and pay for it!
constant, early attacks by white hats, fast response, seem like the only way to build a truly secure product
and now, having lived it, I wouldn't trust anyone who doesn't do that to be secure enough
like every software you use, box.ascii.dev might sometimes be vulnerable to experienced attackers
but unlike most software you use, box will be made as safe as possible, through constant challenge and fixing
bounty's still up, bounties ftw
stay safe
Uri Tzadikevitch retweeted
Hi Mark 👋
Congrats on the release, Muse Spark 1.2 is a great model!
You forgot to test it on a real organizational environment, but no worries, we did it with @boundarybench
Today we are open-sourcing @boundarybench, a new paper, benchmark, and GitHub repo that enterprises can use to find out the REAL performance of their agents.
Boundary-Bench was developed by researchers from @Accomplish_ai and NYU, where we tested 12 frontier agents across roughly 10,000 runs, with realistic enterprise policies, simulating environments with EDR, SASE, and DLP security tools enforcing those policies.
We did this because generic leaderboard scores are being generated under conditions no security team would ever allow, which means orgs are making deployment and risk decisions based on numbers that don't hold up.
The results are surprising >>





























