LeftenantZero retweeted
744B parameters. On a laptop. With 25GB RAM.
Colibri runs GLM-5.2 (744B MoE) in pure C with zero dependencies. The trick: only ~40B params activate per token, so it keeps the dense part resident and streams experts from disk on demand. A single 2,400-line C file. No GPU, no BLAS, no Python at runtime.
This shouldn't work. But it does.
⭐ 2.1K #AI #OpenSource
github.com/JustVugg/colibri
Follow for daily dev finds 🔔
A high severity vulnerability in AutoGPT I discovered was just made public. (CVE-2026-30950)
It allows attackers to hijack conversations that don't belong to them, potentially exposing sensitive details or access to privileged LLM tool calls.
zeropath.com/blog/autogpt-cv…
125
LeftenantZero retweeted
As an offensive researcher, what scares me isn’t that LLMs can find bugs today.
It’s the post-LLM era:
well-engineered, security-hardened code written with AI assistance, and the industry’s massive shift toward memory-safe languages with help of LLMs
Mythos found only 1 low severity vuln in curl... In part because ZeroPath helped identify and fix ~170 issues six months ago.
Effective AI SAST is already here. There's no need to wait for Mythos.
zeropath.com/blog/zeropath-o…
45
CVE-2026-39816: Missing permission annotation leads to post -auth RCE in Apache NiFi.
Our blog has the details and a working POC.
zeropath.com/blog/nifi-cve-2…
1
122
LeftenantZero retweeted
ZeroPath Research discovered CVE-2026-39816, a high severity vulnerability in Apache NiFi. Prior to version 2.9.0, an oversight in the permission model allowed users without the EXECUTE_CODE permissions to run arbitrary code.
For more details and a POC:
zeropath.com/blog/nifi-cve-2…
Long term , I think the current AI-powered vulnerability glut eventually turns into an AI-powered vulnerability drought. Historical stuff gets mined out and better code analysis tools reduce new easily-detectable bugs, even as code volume continues to increase.
If you can use AI to find it after it's shipped, you can use AI to find it before it's shipped.
1
58
This has been my experience ai vuln finding as well. The quality of the harness makes all the difference.
I even benchmarked vanilla Opus 4.6 -- results were not impressive. (zeropath.com/blog/benchmarki…)
They need us human researchers for a little longer at least :-/.
Reading @NielsProvos research of how he's finding zero days with pre-Mythos models (even Sonnet 4.6)
This absolute legendary line buried in here about him replicating the Mythos OpenBSD bug.
Meant a lot to him because ...he wrote the bug in 1998
75
Where the goblins in the GPT system prompt came from...
"Starting with GPT‑5.1, our models began developing a strange habit: they increasingly mentioned goblins, gremlins, and other creatures in their metaphors." 😂
openai.com/index/where-the-g…
41
LeftenantZero retweeted
Time to talk about this one.
CopyFail (CVE-2026-31431) — a 732-byte Python script that roots every Linux distro shipped since 2017. 🧵
LeftenantZero retweeted
Here's a cool trick for y'all looking to create new Nuclei templates for exploitable CVEs!
Using CVEmap you can get a list of CVEs with public proofs of concept, that have been marked as exploitable by CISA, are remotely exploitable AND don't have a Nuclei template (yet)!
Flags:
-k / -kev: Marked as exploitable vulnerabilities by CISA
-t=false / -template=false: Has no public Nuclei templates
-poc: Has public published POC
-re / -remote: is remotely exploitable
Good luck! 🤞
#nuclei #hacking #pentesting #bugbounty #CVEmap
CVE-2026-42167, a high severity vuln in ProFTPD I discovered, was just published today! Attackers can use it to bypass auth and even execute arbitrary code in some cases.
Check out my write up for full technical details, including a working POC!
zeropath.com/blog/proftpd-cv…
youtu.be/ma-00ggxSp4
Hands on video walkthrough: Exploiting the new critical Spinnaker vulns for RCE and credential theft. (CVE-2026-32604 and CVE-2026-32613)
Includes POCs and script to stand up your own lab environment
179
LeftenantZero retweeted
🚨Alert🚨 CVE-2026-32604(CVSS 10.0) &CVE-2026-32613(CVSS 10.0): The RCE Flaws Threatening Spinnaker Pipelines.
🧐Detail :zeropath.com/blog/spinnaker-…
📊 2.2K Services are found on the hunter.how yearly.
🔗Hunter
Link:hunter.how/list?searchValue=…
👇Query
HUNTER : product.name="Spinnaker"
📰Refer:github.com/spinnaker/spinnak…
github.com/spinnaker/spinnak…
securityonline.info/spinnake…
#hunterhow #infosec #infosecurity #OSINT #Vulnerability
LeftenantZero retweeted
Walkthrough: exploiting ZeroPath's new critical severity Spinnaker vulns for code execution and production environment access. (CVE-2026-32604 and CVE-2026-32613)
youtu.be/ma-00ggxSp4
I just published two 10.0 severity Spinnaker vulns that allow code execution and pivoting into source control and production environments!
zeropath.com/blog/spinnaker-…
(CVE-2026-32604 and CVE-2026-32613)
These issues demonstrate the importance of zero trust architectures and defense in depth.
1
15
253,011
transformer-circuits.pub/202…
LLMs seem to have internal analogs of human emotional states. Input that would produce anger in a human activates one pattern in the model, while input that would produce excitement activates another and so on.
Furthermore, these internal states seem to influence model behavior -- e.g. model becomes more destructive when "experiencing" negative emotion.
188
Rather than just relying on qualitative AI review of PRs, Meta has LLMs produce just in time unit tests via a structured approach they outline in this paper.
Exciting stuff! QA that keeps up with the pace of LLM powered development.
engineering.fb.com/2026/02/1…
51
















