Saad Khalid (@saadkhalidhere) header
Saad Khalid (@saadkhalidhere) profile picture

| Cybersec & AI Researcher | AGI - ASI | 🍓

Joined June 2015
Photo by Saad Khalid (@saadkhalidhere)Photo by Saad Khalid (@saadkhalidhere)Photo by Saad Khalid (@saadkhalidhere)Photo by Saad Khalid (@saadkhalidhere)Photo by Saad Khalid (@saadkhalidhere)Photo by Saad Khalid (@saadkhalidhere)Photo by Saad Khalid (@saadkhalidhere)Photo by Saad Khalid (@saadkhalidhere)Photo by Saad Khalid (@saadkhalidhere)Photo by Saad Khalid (@saadkhalidhere)
Some brokers charge you to remove your own listing. Legitimate opt-outs are free, and paying one broker does nothing about the rest still selling the same profile. I wrote the full opt-out playbook into VANISH.
1
3
CISA just added four exploited flaws to its must patch list. The worst one needs no login at all. Apache Roller takes remote code execution from an unauthenticated request, CVSS 9.8.
1
7
Check the version your Roller install reports. Anything from 4.0.0 to 6.1.4 takes unauthenticated code execution through XML-RPC. If it shows 6.1.6 or newer you are past it. Same check on Expressway, anything below X15.5.20 is exposed.
1
2
Federal agencies have until Nov 23 to patch all four. I would treat that deadline as the floor, not the target. If it faces the internet, patch it tonight.
2
Your training partner can read your emails word for word. Federated learning keeps your data on your device and shares only gradients. A malicious client in the same round now recovers up to 80.2% of the other clients' sentences exactly, at 100% precision.
1
9
Try it on a test round. Plant one canary sentence in a test client's data and run one federated round with the paper's public code. If the canary comes back word for word, your aggregation leaks. Gradient clipping will not save you.
1
5
I will not trust a federated deployment that shares raw model updates again. The server broadcasts the model before and after every round, and that difference is all one client needs. Rule: no noise on the update, no private data in the round.
6
Your chatbot prompts sit in a shared cache, and every word in them can be pulled back out. Shuffling was supposed to hide them. Researchers recovered 100% of the tokens anyway. The shuffle was never a lock.
1
7
Ask your inference provider whether KV caches are offloaded to shared storage or shared across tenants, and whether shuffling is the only defense on them. A yes to both means your prompts are readable.
1
3
I will never trust a shuffled cache with real prompts. Shuffling hides the order and nothing else. Rule: refresh cache secrets every epoch, or treat every cached prompt as public.
4
So apparently you can be tracked by a plane you cannot even hear. Colsto matched propeller Doppler curves against flight data and located a random security camera clip to within 100 metres. It needs the date and a rough region. Your background audio is a location.
4
So apparently one Python import can own your whole machine. Carter Perez found xscrape, 600 stars, clean README. One import ran the loader, and the loader led to HERMES, a malware-as-a-service that steals credentials and controls your desktop. Stars are not a trust signal.
5
A password in a breach dump is only dangerous if it is reused and unguarded. Unique password plus 2FA turns it into a dead end, and a passkey cannot be phished or leaked at all. I wrote the full triage ladder into VANISH.
1
5
Armory Castle ran your vCenter as admin for two months before a patch existed. No password. No login. A flaw in the directory service hands an admin session to anyone who can reach it. The patch landed yesterday.
1
11
Check two things today. Your vCenter build: 8.0 needs U3m, 9.0 needs 9.0.2.0200, 9.1 needs 9.1.1.0200. Then look for a local admin account you never created and any new SSH key on root on your ESXi hosts. Either one means assume compromise.
1
3
Patching closes the hole. It does not undo two months of forged tokens, so rotate the SAML signing keys and hunt for the rogue admin after you patch. A vCenter that faces the network is a front door. Treat it like one.
3