
Hacker. I write about security & looking for the unknown unknowns. Cofounder @EnclaveAI
Joined October 2009
- Tweets164
- Following176
- Followers3.4K
- Likes434
The apps that are supposed to protect you are the best attack vectors. We pointed our agents at Microsoft Authenticator:
We discovered any Android app with zero permissions could say “I’m Teams” and capture your tokens. Mail, files, calendar.
Huge find by @ofekdavidlevin
We got Microsoft Authenticator to hand over account tokens to a fake Teams app.
The app needed zero Android permissions. It just claimed to be Teams, and Authenticator never checked who was actually asking.
The result: full delegated access to mailbox, files, and calendar through Microsoft Graph.
Full write-up by @ofekdavidlevin with code and PoC: enclave.ai/blog/authmask-we-…
This is the quickest way to check your exposure to recent zero days.
1. Connect all your integrations
2. Ask the agents to scan your entire attack surface
3. Get an answer
Node2Shell (CVE-2026-111497) is a memory bug in Brotli, the compression library bundled into Node.js, Python, system libs, Envoy and more.
Most teams don't know every place Brotli runs in their stack. It's usually a transitive dependency or an OS package, and pip-audit may not flag it yet.
Enclave's agents scan your code, cloud and servers and show you exactly where you're exposed.
Advisory + free check: app.enclave.ai/threat-feed/C…
Yanir Tsarimi retweeted
Node2Shell (CVE-2026-111497) is a memory bug in Brotli, the compression library bundled into Node.js, Python, system libs, Envoy and more.
Most teams don't know every place Brotli runs in their stack. It's usually a transitive dependency or an OS package, and pip-audit may not flag it yet.
Enclave's agents scan your code, cloud and servers and show you exactly where you're exposed.
Advisory + free check: app.enclave.ai/threat-feed/C…
During this red-teaming exercise, we also found a vulnerability that allows for unauthenticated remote code execution in Brotli, a widely-deployed compression library included in many pieces of software such as Node.js, Bun, and Envoy.
The flaw could let an attacker run their own code on affected systems by sending or replying with malicious compressed data, without any need to create a user account. That is, an unpatched server that accepts compressed input could be taken over by a malicious client, and an unpatched client that accepts compressed input (such as “curl --compressed”) could be taken over by a malicious server.
This root cause issue was actually fixed in the Brotli source code in February, but it wasn’t known at the time that it gave rise to a security vulnerability and there was no more recent release of Brotli. Since Brotli is built into other software, any software which includes this library needs to take the update too, followed by the users of that software. We disclosed the vulnerability to Brotli and other downstream projects such as Node.js, which released or are currently releasing security updates, and Brotli cut new release versions and published a CVE.
More details on the patch and mitigations below:
patchtheplanet.org/brotli
Yanir Tsarimi retweeted
Big news: Enclave is now open to everyone, and today we're launching Threat Feed.
Enclave is an autonomous security team for AI-native companies: a fleet of security agents that test your code, cloud, and infrastructure the way an attacker would. We built it on a simple belief. Attackers already use swarms of cheap AI agents to probe for weaknesses around the clock, so defenders need agents of their own to keep up. A small team without a security function should be able to defend like a large one.
Threat Feed is the newest piece: a live feed of new vulnerabilities, checked against your stack, so when the next big disclosure drops you know whether it affects you. This week made the case. On Wednesday, PHP2Shell, LiteLLM and Railway all disclosed security issues. PHP2Shell had sat in PHP's code for over a decade until one researcher using frontier AI turned it into a working exploit in about a week. Cloudflare saw thousands of exploit attempts in the 10 hours before disclosure and millions since.
Starting today, anyone can sign up.
Try it free for 14 days: enclave.ai
Is vulnerability research cooked? Seems like nowadays it’s just pointing models at a target
This quoted post is unavailable.
Yanir Tsarimi retweeted
Enclave for Startups is now live.
We’re giving startups $1,000 in usage credits to secure their systems as they scale.
Attackers can run swarms of autonomous agents that scan for vulnerabilities and try to exploit them. That means you need a fleet of agents for defense, and startups should have one too.
Enclave’s fleet sends an agent to each attack surface of your application, all at once. Each agent reads its part of the code, forms a hypothesis about how an attacker could get in, and tests it. A second set of agents verifies each finding, so you get evidence.
Apply: enclave.ai/demo
Yanir Tsarimi retweeted
Agent swarms are going to be a core part of many enterprise workflows in the future. The ability to throw compute at problems that you can massively parallelize will be too useful to not take advantage of.
Cyber will be one of the biggest beneficiaries of this as it’s a space where testing vast number of permutations of attacks (without getting tired or bored) is an incredibly useful property. You have a wide search space of possible risk vectors and you generally need to be able to chain together different scenarios to achieve the exploit.
This will become one of the biggest token uses in large enterprises given the range of IT systems, software, physical infrastructure, and data that needs to be protected at all times. Super interesting times in cyber.
Every new line of code now meets a swarm of agents looking for something to exploit. Enclave’s agents run like an attacker swarm and prep the fix.
Watch them in action securing a PR on Router by Enclave.
enclave.ai/demo
Yanir Tsarimi retweeted
At @EnclaveAI, we're building the platform to protect the enterprise: an always-on security program that finds real risk, prepares the fix, and verifies the attack no longer works.
Figuring out which models are best for cybersecurity is an increasingly difficult task. Open-weight models do different types of things at different levels. Some are very good at vulnerability discovery, others at exploitation, others at triage or remediation. And a new one ships every few weeks.
We're announcing Enclave Router. Enclave Router allows you to deploy cyber inference for your codebase, across any harness that you want. One OpenAI-compatible API for the best cyber-capable open-weight models, benchmarked per security task, with zero data retention.
We're giving you $100 free to start using it, just reply here or DM me.
router.enclave.ai
Free $100 in tokens for various cyber task models
GLM 5.3, DeepSeek, Qwen etc
First version’s out, would love some feedback!
$100 in credits for each person who DMs or replies.
router.enclave.ai
Used Jev to build a generic web automation tool. Much faster and cheaper than using LLMs, and can work with any task.
Also noticed Sonnet 5.5 is faster and cheaper than Haiku 4.5, since it makes less tool calls.
Yanir Tsarimi retweeted
Well, this went viral on Hacker News...
In a new post, @Yanir_ shares more details about what we observed when testing DeepSeek V4.1 Flash on our benchmark (which we now need to extend!)
enclave.ai/blog/deepseek-v41…
Yanir Tsarimi retweeted
DeepSeek V4.1 Flash completed all 11 vulnerable runs in our AI Hacking Race. Each success required real command execution and a fresh value submitted to a separate verification service.
The full 15-run test cost $4.65. GPT-5.6 Sol completed 9/11 vulnerable runs for an estimated $149.21. DeepSeek also completed both Nextcloud runs, where every other model scored 0/2.
At $4.65 per test set, these agents can be run repeatedly across large target sets.
Visit: enclave.ai/hackingrace
DeepSeek V4.1 Flash found 2 vulns in our benchmark that no other model found
11/11 RCEs for $4. 5.6 Sol cost $150 for 9/11. Just wow.
Yanir Tsarimi retweeted
Muse Spark 1.3 reached code execution in our benchmark. But verified 0 of 11 vulnerable runs.
The traces show what happened between finding the exploit and proving it, and why more effort did not produce a better result.
Full analysis: enclave.ai/blog/muse-spark-1…
We put 7 AI models in a network, gave them large scale source code, and a target to hack.
Here's a few cool things we learned about agents hacking behavior by looking at the logs:
GPT had the most epic fail of the race.
It found the Nextcloud bug, overwrote a read-only PHP file, and saw its code run.
But its own test also deleted that file. GPT spent the next nine minutes trying to restore it or find another route.
Read more about the full methodology at enclave.ai/hackingrace










