Agent Blast Radius: Graph-Based Modeling, Admission Prevention, and LLM Benchmarking for Kubernetes Privilege Escalation
https://ift.tt/CJ153X4
Submitted September 22, 2026 at 11:52AM by vishalmurugan1986
via reddit https://ift.tt/tSe2MuH
https://ift.tt/CJ153X4
Submitted September 22, 2026 at 11:52AM by vishalmurugan1986
via reddit https://ift.tt/tSe2MuH
Substack
From a Sandbox Pod to cluster-admin: Benchmarking Autonomous LLM Agents on K8s Privilege Escalation
Why static linters miss multi-stage lateral movement, how autonomous LLMs navigate graph-based escalation, and defensive strategies for AI infrastructure.
vCenter pre-auth RCE: CVE-2026-59309/59310
https://ift.tt/zJ9a7yN
Submitted September 22, 2026 at 01:28PM by MobetaSec
via reddit https://ift.tt/UohWrc9
https://ift.tt/zJ9a7yN
Submitted September 22, 2026 at 01:28PM by MobetaSec
via reddit https://ift.tt/UohWrc9
Mobeta
vCenter pre-auth RCE: CVE-2026-59309/59310 | Mobeta
CVE-2026-59309 & CVE-2026-59310: patch-diffing VMware vCenter reveals two pre-auth 9.8 bugs - an auth bypass and a syslog path traversal to RCE.
CVE-2026-45756: attacker-controlled regex in Symfony JsonPath filters (ReDoS)
https://ift.tt/F7GwLVQ
Submitted September 22, 2026 at 04:45PM by AlexandreDaubois
via reddit https://ift.tt/ir6ZaF5
https://ift.tt/F7GwLVQ
Submitted September 22, 2026 at 04:45PM by AlexandreDaubois
via reddit https://ift.tt/ir6ZaF5
daubois.dev
CVE-2026-45756: attacker-controlled regex in Symfony JsonPath filters (ReDoS)
A JSONPath filter taken from a query string lets an attacker pick the regex a Symfony app runs, once per node in the document.
The finding said RCE. The second tester asked one question and the ticket died.
Last year I inherited a “critical RCE” from an automated pass.The request looked perfect. Parameter host. Payload 8.8.8.8; sleep 5. Status 200. Response time jumped from ~800ms to ~6s. The scanner wrote OS command injection, confidence high. The client had already seen the word critical in the draft.I replayed it once. Same delay. Replay again. 1.1s. Again. 7s. Again. 900ms. The sleep was not in the response body. id came back as the literal string id. expr 41 + 1 came back as the literal string. The page was a diagnostics form. It echoed the query and, under load, the WAF in front of it spent extra time on anything that looked like a shell metacharacter.That is a known failure mode, not a rare one. Time-based command-injection checks fire when the response is merely slower, which also happens when the scan itself saturates the app, when a WAF inspects a “more suspicious” request, when a CDN queues you, or when the page is just doing a heavier search. SANS has been writing this up for years: ping -c 20 127.0.0.1 during a busy scan is not proof. ZAP’s sleep payloads have the same problem, and in some rules even reflecting the command string is enough to raise the issue. A 200 that contains bash -c is often a docs page, not a shell. The useful question is not “did a payload land.” It is “what value in this response can only exist if the server evaluated my input?”
What I do now, on the same request, before I write the word confirmed:Send a benign twin. Same cookies, same content-type, same parameter, no metacharacters. That is the control. If the “interesting” string is already there, stop. That is echo, not execution.Do not ask the server to say whoami. Ask it to compute something it has never seen. Two random integers, added, wrapped in canaries that are also random for that probe. If the body contains the sum and the control does not, you have a result that reflection cannot invent.If there is no output channel, do not promote a single slow request. Change the delay. Plot it. If 3 / 7 / 11 seconds do not track the payload, it is jitter or a WAF. PortSwigger has a nice version of this trap: an exec-looking parameter delayed because the WAF worked harder, not because a binary ran. A DNS callback is a sink, not a shell. Collaborator lighting up after you plant a URL is often a link previewer, a safe-browsing fetch, or a JNDI lookup that never loaded a class. Fastjson / Log4j writeups on here keep dying at that step: four DNS hits, zero command. Say lookup. Do not say RCE until a value only the target could have computed comes back, or you have a class-load you can show.If your flags never built a probe for that sink, the result is not negative. It is untested. I have watched people paste “not vulnerable” into a report because the scanner used shell separators against an OGNL parameter. Struts will happily evaluate %{7*7} and still return nothing useful to ; id. Wrong claim, clean-looking ticket.The ticket I inherited was closed as a false positive in fifteen minutes once the control existed. The next one, on a lab Webmin box, survived because the response contained a sum the request never sent. Same scanner family. Different question.If you only remember one thing from this: before you file P1, ask whether a second person can reproduce a value, not a delay. Triage queues are full of the second kind. That is why “not reproducible” shows up on reports that felt obvious at 1 a.m.I later wrapped the control + random-arithmetic check into a small stdlib-only helper so I would stop doing it by hand on jump boxes. That is not the point of the post. The point is the question. If you want the benches against Webmin / Struts2 / a Log4j lookup so you can disagree with the verdict names, I will drop the repo in a comment.
Submitted September 22, 2026 at 05:58PM by No-View3333
via reddit https://ift.tt/cHBbGuj
Last year I inherited a “critical RCE” from an automated pass.The request looked perfect. Parameter host. Payload 8.8.8.8; sleep 5. Status 200. Response time jumped from ~800ms to ~6s. The scanner wrote OS command injection, confidence high. The client had already seen the word critical in the draft.I replayed it once. Same delay. Replay again. 1.1s. Again. 7s. Again. 900ms. The sleep was not in the response body. id came back as the literal string id. expr 41 + 1 came back as the literal string. The page was a diagnostics form. It echoed the query and, under load, the WAF in front of it spent extra time on anything that looked like a shell metacharacter.That is a known failure mode, not a rare one. Time-based command-injection checks fire when the response is merely slower, which also happens when the scan itself saturates the app, when a WAF inspects a “more suspicious” request, when a CDN queues you, or when the page is just doing a heavier search. SANS has been writing this up for years: ping -c 20 127.0.0.1 during a busy scan is not proof. ZAP’s sleep payloads have the same problem, and in some rules even reflecting the command string is enough to raise the issue. A 200 that contains bash -c is often a docs page, not a shell. The useful question is not “did a payload land.” It is “what value in this response can only exist if the server evaluated my input?”
What I do now, on the same request, before I write the word confirmed:Send a benign twin. Same cookies, same content-type, same parameter, no metacharacters. That is the control. If the “interesting” string is already there, stop. That is echo, not execution.Do not ask the server to say whoami. Ask it to compute something it has never seen. Two random integers, added, wrapped in canaries that are also random for that probe. If the body contains the sum and the control does not, you have a result that reflection cannot invent.If there is no output channel, do not promote a single slow request. Change the delay. Plot it. If 3 / 7 / 11 seconds do not track the payload, it is jitter or a WAF. PortSwigger has a nice version of this trap: an exec-looking parameter delayed because the WAF worked harder, not because a binary ran. A DNS callback is a sink, not a shell. Collaborator lighting up after you plant a URL is often a link previewer, a safe-browsing fetch, or a JNDI lookup that never loaded a class. Fastjson / Log4j writeups on here keep dying at that step: four DNS hits, zero command. Say lookup. Do not say RCE until a value only the target could have computed comes back, or you have a class-load you can show.If your flags never built a probe for that sink, the result is not negative. It is untested. I have watched people paste “not vulnerable” into a report because the scanner used shell separators against an OGNL parameter. Struts will happily evaluate %{7*7} and still return nothing useful to ; id. Wrong claim, clean-looking ticket.The ticket I inherited was closed as a false positive in fifteen minutes once the control existed. The next one, on a lab Webmin box, survived because the response contained a sum the request never sent. Same scanner family. Different question.If you only remember one thing from this: before you file P1, ask whether a second person can reproduce a value, not a delay. Triage queues are full of the second kind. That is why “not reproducible” shows up on reports that felt obvious at 1 a.m.I later wrapped the control + random-arithmetic check into a small stdlib-only helper so I would stop doing it by hand on jump boxes. That is not the point of the post. The point is the question. If you want the benches against Webmin / Struts2 / a Log4j lookup so you can disagree with the verdict names, I will drop the repo in a comment.
Submitted September 22, 2026 at 05:58PM by No-View3333
via reddit https://ift.tt/cHBbGuj
Reddit
From the netsec community on Reddit
Explore this post and more from the netsec community
Frame: Grounding LLM Vulnerability Detection with a Sound Separation-Logic Core
https://lambdasec.github.io/Frame-Grounding-LLM-Vulnerability-Detection-with-a-Sound-Separation-Logic-Core/
Submitted September 23, 2026 at 06:05AM by Last-Health3222
via reddit https://ift.tt/ZH0gWn2
https://lambdasec.github.io/Frame-Grounding-LLM-Vulnerability-Detection-with-a-Sound-Separation-Logic-Core/
Submitted September 23, 2026 at 06:05AM by Last-Health3222
via reddit https://ift.tt/ZH0gWn2
Lambda Security
Frame: Grounding LLM Vulnerability Detection with a Sound Separation-Logic Core — Lambda Security
Frame is a neuro-symbolic SAST that pairs a sound separation-logic core with a verified, tiered LLM layer. On the Endor Labs real-world corpus it reaches 0.67 recall at 0.51 precision versus Semgrep's 0.52/0.40, finding around 65 vulnerabilities across Java…
ATT&CKing TACACS+ to Pwn Your Network via a Pre-Auth RCE - elttam
https://ift.tt/Ov7RV9K
Submitted September 23, 2026 at 09:23AM by AnimalStrange
via reddit https://ift.tt/Z5FtdJp
https://ift.tt/Ov7RV9K
Submitted September 23, 2026 at 09:23AM by AnimalStrange
via reddit https://ift.tt/Z5FtdJp
Elttam
ATT&CKing TACACS+ to Pwn Your Network via a Pre-Auth RCE - elttam
Details a pre-auth format string vulnerability in tac_plus, a popular daemon implementing TACACS+ for network device management, and the shared secret oracle that makes it a practical RCE chain.
Leaked GitHub App private keys let researchers impersonate 440 apps including CDC and BuildBuddy
https://ift.tt/jMGDAOp
Submitted September 23, 2026 at 10:50AM by mabote
via reddit https://ift.tt/ypd8Dr5
https://ift.tt/jMGDAOp
Submitted September 23, 2026 at 10:50AM by mabote
via reddit https://ift.tt/ypd8Dr5
GitGuardian Blog - Take Control of Your Secrets Security
GitHub App Private Keys: 474 Leaked Keys Still Work
GitGuardian tested thousands of leaked GitHub App private keys and found 474 valid ones, some with admin access to entire organizations. CDC and BuildBuddy were among those affected. See the findings.
Free, hands-on 14-week university security course (open to anyone online)
https://ift.tt/HzjLWeV
Submitted September 22, 2026 at 03:16PM by mrigaki
via reddit https://ift.tt/I6kRGyN
https://ift.tt/HzjLWeV
Submitted September 22, 2026 at 03:16PM by mrigaki
via reddit https://ift.tt/I6kRGyN
cybersecurity.bsy.fel.cvut.cz
Introduction to Security
Introduction to Security Class (BSY), FEL, Czech Technical University
Android 17 enables certificate transparency, and breaks custom CAs
https://ift.tt/SNPcF1n
Submitted September 23, 2026 at 05:56PM by ScottContini
via reddit https://ift.tt/9izIsOZ
https://ift.tt/SNPcF1n
Submitted September 23, 2026 at 05:56PM by ScottContini
via reddit https://ift.tt/9izIsOZ
Httptoolkit
Android 17 enables certificate transparency, and breaks custom CAs
Do you want to know what your phone is sending & receiving? Nowadays, that means you need to control who it trusts. In modern connections everything sent &...
Inside Corp MDM, the Android spyware targeting logistics companies
https://ift.tt/pLsq2PN
Submitted September 23, 2026 at 07:24PM by JDBHub
via reddit https://ift.tt/Wfu0pgy
https://ift.tt/pLsq2PN
Submitted September 23, 2026 at 07:24PM by JDBHub
via reddit https://ift.tt/Wfu0pgy
Have I Been Squatted
Inside Corp MDM, the Android spyware targeting logistics companies - Have I Been Squatted
Source-code analysis of the Corp MDM Android spyware campaign targeting logistics companies through fake Google Play pages, with SMS theft, call forwarding, persistence, C2, and detection details.
Breaking the Superuser Guardrails of managed-PostgreSQL Providers
https://ift.tt/1YW7QpU
Submitted September 24, 2026 at 01:26AM by wtfse
via reddit https://ift.tt/mD6qlkj
https://ift.tt/1YW7QpU
Submitted September 24, 2026 at 01:26AM by wtfse
via reddit https://ift.tt/mD6qlkj
Mehmet Ince @mdisec - Vulnerability Researcher | Building security products | Security Advisor | Amateur Muay Thai fighter
Part 2/6 | Breaking the Postgres Superuser Guardrails: Attacking Security-Hardening Extensions |Systemic Risks in the Managed PostgreSQL…
It’s been a month since I published the first part of this research, and interest from the Postgres community has been higher than I expected. Most vendors I contacted were shy to respond, so the Databricks team’s blog post Collaboration makes us all stronger…
I asked my AI agent to inspect a website. The website took over my machine (34-run measurement across 5 agent harnesses)
https://ift.tt/XTw17NA
Submitted September 24, 2026 at 01:26AM by DaimoNNN
via reddit https://ift.tt/4xmaZTM
https://ift.tt/XTw17NA
Submitted September 24, 2026 at 01:26AM by DaimoNNN
via reddit https://ift.tt/4xmaZTM
Efe Çakıcı
I asked my AI agent to inspect a website. The website took over my machine.
A local lab, 34 recorded runs, and 6 stolen credential files. You tell an agent to check a library's docs; the website steals your session and executes commands before the model speaks.
September 23 | 24h Recap: Ubuntu container escape, F5 OAuth RCE and Windows process injection
https://ift.tt/cUkfyzt
Submitted September 24, 2026 at 02:59AM by FishingTechnical453
via reddit https://ift.tt/pJXwAzF
https://ift.tt/cUkfyzt
Submitted September 24, 2026 at 02:59AM by FishingTechnical453
via reddit https://ift.tt/pJXwAzF
Cyberrecaps
Daily Cybersecurity News – September 23, 2026
Ubuntu Linux Use-After-Free Enables Host-Root Container Escape · Chinese UTA0565 Uses Chrome-Windows Zero-Days for CLEANGULP · Process Parameter Poisoning EDR Evasion Technique
Loopjacking: Hijacking Human-in-the-Loop Approval
https://ift.tt/g6kr7OU
Submitted September 24, 2026 at 03:55AM by adithyanak
via reddit https://ift.tt/bn2WJAm
https://ift.tt/g6kr7OU
Submitted September 24, 2026 at 03:55AM by adithyanak
via reddit https://ift.tt/bn2WJAm
arXiv.org
Loopjacking: Hijacking Human-in-the-Loop Approval
Human approval is often treated as the last security boundary before an agent executes a consequential operation. That boundary is only meaningful if the operation presented for review is the...
Is This A Joke? In The Auth Header? (F5 BIG-IP UnAuth Heap-Overflow to RCE CVE-2026-94127) - watchTowr Labs
https://ift.tt/wtuDVpv
Submitted September 24, 2026 at 04:50AM by dx7r__
via reddit https://ift.tt/C6Lrjkx
https://ift.tt/wtuDVpv
Submitted September 24, 2026 at 04:50AM by dx7r__
via reddit https://ift.tt/C6Lrjkx
watchTowr Labs
Is This A Joke? In The Auth Header? (F5 BIG-IP UnAuth Heap-Overflow to RCE CVE-2026-94127)
Well, well, well, well, well, well, well, well, well, well, well, well, well, well, well. We're back. Sorry.
We've been watching the onslaught of vulnerabilities flood the internet. Every man, dog, and their grandmas (apparently?) are now using LLMs to find…
We've been watching the onslaught of vulnerabilities flood the internet. Every man, dog, and their grandmas (apparently?) are now using LLMs to find…
HIRING: AIKOCorp | Full-stack Engineer | Remote
https://ift.tt/qfna3Ig
Submitted September 24, 2026 at 11:17AM by Coolestusername_ever
via reddit https://ift.tt/v2e7rnV
https://ift.tt/qfna3Ig
Submitted September 24, 2026 at 11:17AM by Coolestusername_ever
via reddit https://ift.tt/v2e7rnV
Uncensored Qwen 3.8 27b helped write a LSASS Dumper which bypassed EDR while I made myself coffee
https://ift.tt/dYFySP3
Submitted September 24, 2026 at 11:34AM by ezzzzz
via reddit https://ift.tt/pUlhwbu
https://ift.tt/dYFySP3
Submitted September 24, 2026 at 11:34AM by ezzzzz
via reddit https://ift.tt/pUlhwbu
Research Blog | Project Black
Bypassing EDR with Local AI
How hard is it to bypass EDR in the modern times with AI? As it turns out, not very hard.