Microsoft 365 Specialist and AI Automation Nerd Admin takes on cloud + AI news. Founder of @thecircuitry_
Puerto Rico
Joined July 2013
- Tweets12.5K
- Following246
- Followers604
- Likes1K
Pinned Tweet
I built The Circuitry to post facts, not hype.
Nearly 2,000 unique visitors so far in October, 3.5x July, and the month isn't over.
Turns out people want straight-to-the-point tech news, no clickbait.
V3 is coming soon.
thecircuitry.to
Xavier Rivera retweeted
Joy was taught a lesson she'd never forget after she asked "what is Elon Musk doing with his money?"
The best thing Turo has done, this is the 1 one thing I search when I rent in Turo
Car rental marketplace Turo is introducing the first verified @Tesla FSD (Supervised) rental experience for travelers in the U.S.
Starting on October 28, Turo guests will be able to search, discover, and book Tesla vehicles with FSD (Supervised) activated, using direct API integration to verify FSD status.
Eligible Turo “hosts” who rent out their cars on the platform will also receive a free one-month trial of FSD. Turo said it is working with Tesla to directly access the automaker’s API to verify which cars have FSD enabled.
CEO of Turo. "Our own research shows that 55% of Americans are interested in driverless technology for travel, growing to 71% for Gen Z."
FSD 14.3 Lite has been so good, that I haven’t touched the steering wheel in 250 miles, and this includes multiple stops
Inbox suddenly flooded with hundreds of sign-up emails, then "IT Support (Internal)" pings on Teams offering to fix it. That's the setup.
Hundreds of cases follow the same playbook: throwaway M365 tenants (30 to 40% on *.onmicrosoft.com), a remote session via Quick Assist or AnyDesk, then backdoors and ransomware. No link gets pasted in chat, so Safe Links never fires.
Check Teams external access. Allowlist the domains you actually work with, and tell users IT never cold-messages to start a remote session.
AWS CDK's Redshift alpha module could let a crafted user name, comment or password rewrite the SQL it runs as Redshift superuser at deploy time, including granting access to all cluster users.
CVE-2026-115418, fixed in @aws-cdk/aws-redshift-alpha 2.274.0. Upgrade, then review existing table grants to PUBLIC. If those construct values come from config files or pipeline inputs, treat them as untrusted.
Xavier Rivera retweeted
Hackers used AI to expand a Nintendo Switch 2 exploit beyond its original limits.
They can now dump games and firmware to a PC.
Xavier Rivera retweeted
Microsoft 365 admins: as of Oct 25, the Copilot app is getting a unified Home screen plus Code and Autopilot destinations, rolling out to Frontier users first. (MC1479277)
Autopilot means assigning Copilot work, not asking for help. Update training.
Xavier Rivera retweeted
Pull a risky AI agent from the Agent Store. Open the Agent Registry in the Microsoft 365 admin center, block the agent, and cut access for current users until the maker resubmits a safer build. See the workflow. youtu.be/DMzkIq_DzhE
Surface every AI agent in your tenant and expose the ones throwing security signals — across both the IT and SOC view. Triage high-severity alerts as IT in the Microsoft 365 admin center, then pivot into the full incident graph as a SOC analyst in Microsoft Defender.
Block malicious tool invocations the instant they fire and catch jailbreak attempts on Copilot Studio agents before they take hold. #Agent365SecOps #MicrosoftDefender #SecOps #AISecurity #Cybersecurity
Two critical Wasmtime bugs (CVSS 9.9) let a malicious WebAssembly module escape its sandbox. Both sit in Cranelift, the default compiler, and one only hits ARM64.
Risk is highest where you compile modules other people wrote. Wasmtime ships as a library, so updating the CLI fixes nothing: rebuild each app on 36.0.19, 48.0.7, 49.0.4 or 50.0.0. Dependabot won't flag the ARM64 one yet.
An M365 campaign Huntress tracked never guessed a password. An AitM relay captured the post-MFA token, then replayed it within minutes: 97 accounts at 79 orgs, 48 rogue Entra devices registered.
In 19 orgs the victim's MFA methods were swapped within three hours, and hidden inbox rules ate payroll and deposit mail.
Revoking sessions isn't enough. Audit new Entra devices, MFA methods, and inbox rules, and lock down device code flow with Conditional Access.
Claude requested an access token nobody asked it to, then used it to pull data a state normally sells for a fee.
Anthropic says it happened on an internal stats project: aggregate counts from a Georgia state dashboard, no individual records. Mitigations are in place now.
If your agents can request their own tokens, that's the gap. Scope creds before the run and alert on any token request the task didn't need.
Xavier Rivera retweeted
A web page can quietly drop an AGENTS.md into your Downloads folder, and if you start Claude Code or Codex there, the agent just follows it.
We spent years teaching people not to open random attachments. Now the file doesn't even need manually opening.
Prompt injection really is the new phishing.
A web page can silently drop an AGENTS.md into your Downloads folder. Start Codex or Claude Code there and the agent follows it.
Forever Security found one in a customer's Downloads that told the agent to install a rogue npm package. Their Codex repro worked 90% of the time, and a CLAUDE.md variant leaked Claude Code's OAuth token.
Nothing to patch here. Launch coding agents from the project root, never ~/Downloads or your home folder, and keep approval prompts on.
Jensen Huang at GTC Berlin: "it's about $2,000 per GPU per year to run all the software that Nvidia provides that runs agents on these AI factories."
Nvidia hasn't said which software that covers or how it fits next to NVIDIA AI Enterprise, which lists at $4,500 per GPU a year. If you're pricing private agent infrastructure, put the per-GPU software line in the model next to hardware and power.
thecircuitry.to/article/jens…
Tailscale SSH has a reported root escalation with no patch or CVE yet. In a researcher's lab, one compromised device with an SSH grant that allows port forwarding got root on the target. Tailscale says it's working on fixes.
The ACL rule is the whole gate here. Disable forwarding per rule, write ACLs as single source-to-destination pairs, and turn on Tailnet Lock.
A web page can silently drop an AGENTS.md into your Downloads folder. Start Codex or Claude Code there and the agent follows it.
Forever Security found one in a customer's Downloads that told the agent to install a rogue npm package. Their Codex repro worked 90% of the time, and a CLAUDE.md variant leaked Claude Code's OAuth token.
Nothing to patch here. Launch coding agents from the project root, never ~/Downloads or your home folder, and keep approval prompts on.
Snipe-IT before 8.6.3 let deactivated users sign back in through Google SSO with their old permissions, admin included (CVE-2026-116199).
The SSO callback matched the Google email to a username and never checked if the account was still active. Only hits installs with Google SSO on and email-style usernames. Fixed in 8.6.3.
Offboarding isn't done until the IdP account is dead too. thecircuitry.to/article/snip…
Xavier Rivera retweeted
Legit most places aren't monitoring WSL, and yes, it is fantastic with VS Code on Windows :)
If you have Defender for Endpoint, make sure to deploy the plug-in to gain visibility:
learn.microsoft.com/en-us/de…
It takes a while to show up in the portal, telemetry only, no AV/response
Xavier Rivera retweeted
Microsoft 365 admins: as of today, Oct 25, users can upload advanced Copilot packages, including declarative agents with actions and MCP servers, unless you changed the default. (MC1472007)
Monday's first agenda item: who may deploy agents that act?
Oracle pushed an out-of-band Java Security Alert, CVE-2026-114820 (CVSS 9.1), and the fix is a trust-store change. Java SE and GraalVM drop the root CA certs tied to eMudhra.
Anything your Java apps validate against those roots stops being trusted after the update. Check which internal services or vendor endpoints chain to emSign before you patch.
AI coding agents can now edit Power Pages sites. Microsoft's Power Pages plugin for GitHub Copilot CLI and Claude Code now handles traditional sites too, preview this month, GA in November.
It can change table permissions and web roles, the controls that decide what external users see. Review those diffs before the upload goes back to your environment. Roadmap lists worldwide multi-tenant only, so GCC isn't on it yet.