AlphaONE Operations

Grateful for the people who make security possible. Happy Labor Day.

An employee was working at their desk when a Run box blinked onto their screen, and a console window flashed for a fraction of a second and vanished.

The keystrokes came from our laptop in the customer lobby, three desks away, through a $30 radio dongle, into the wireless receiver plugged into the back of their tower.

This is mousejacking. During an authorized physical engagement at a client site, we recognized the wireless keyboard and mouse receivers as the vulnerable class, took a seat in the lobby, and, 90 seconds later, were injecting keystrokes into an unlocked workstation that was actively in use.

It is not a Bluetooth attack, nor a USB attack against Windows. The host sees normal keyboard traffic from a device it already trusts. That is why it is not stopped by EDR, USB device allowlisting, or MFA. To every one of those controls, it looks exactly like the person at the desk typing.

The most common myth we correct with clients: "We use Logitech, so we are patched." Logitech ships two wireless platforms. Unifying (the orange icon) is vulnerable and sits on most reception and front desks. Bolt is not. Bastille Networks disclosed this class in 2016, and it still works in 2026, because long hardware lifecycles and backward-compatible pairing mean the vulnerable receivers never left the building.

Full write-up with the CVEs, the three lab demonstrations, the detection scan, and the fixes:
https://a1ops.com/blog/mousejacking-wireless-hi

We started a recent engagement as a normal user on a company's network. By the afternoon we were reading its full database backups, a live AI model key, and its client account records, without exploiting a single application.

The way in was Kerberoasting, which is Kerberos working exactly as designed. A SQL service account had a password set in 2017 and flagged to never expire. We requested a ticket for that account, cracked it offline (the password was "Winter2017!"), and used it to log into the production database server.

Then it stopped being a normal server. It was also a workstation two people used to write code: an outside contractor with an AI coding assistant installed, and an internal IT generalist. On the same drive as 158 GB of backups sat a homegrown app that forwarded company data to a hosted AI model, kept its passwords in a plain text file, and contained a basic security flaw any code reviewer would have caught. Next to it were the client account workbooks, Social Security numbers, and mailing lists.

We did not exploit any of that. We read it. It was a file share.

The firm's stated position was that it had no development team, so it needed no secure coding standard. But it had people shipping production code, and generative AI made that feel like a reasonable afternoon project. The tool did not cause this. The absent rule did. If a contractor and a couple of IT staff are writing production code with an AI assistant, they are your development team, and they need the access model, the review, and the standard you would give a hired engineer.

An IT admin searched for "PuTTY." One sponsored ad later, the machine was fully compromised. No exploit, no phishing email, no zero-day.

The user did the most ordinary thing an administrator does: searched for a tool by name and clicked the first download link. That link was a paid ad pointing at a typosquatted clone of the real PuTTY site. The installer was signed with a certificate for "NEW VISION MARKETING LLC" (later revoked) and deployed the Oyster backdoor. Within the hour it had registered a "Security Updater" scheduled task firing every 3 minutes and had copied the machine's SAM, SECURITY, and SYSTEM hives, the raw material for password cracking.

Two facts worth knowing: online ads accounted for more than 60% of the malware and phishing observed in 2025, overtaking email, and roughly 8,500 SMB users were hit in just the first four months of 2025 by malware disguised as everyday tools. This campaign has been tied to Rhysida ransomware.

The good news: every artifact was loud, and the fixes are neither expensive nor exotic. Source tools from official sites (chiark.greenend.org.uk for PuTTY, never putty.org), verify signatures, filter lookalike domains, and hunt for short-interval scheduled tasks running rundll32 from AppData.

This Independence Day, we're celebrating the freedom to protect what matters most. Your digital independence is worth defending. Happy 4th of July! 🇺🇸 Ready to secure your future? Visit a1ops.com to learn how we can fortify your defenses.

The Apple TV in Conference Room B had been there for eight months before we told the client it was ours.

IT assumed facilities installed it. Facilities assumed IT installed it. Nobody questioned it. And every night at 2 AM, it opened an encrypted tunnel to an IP address in Frankfurt.

On an authorized physical engagement, our team walked in wearing a maintenance uniform and left four devices behind: the Apple TV running a mesh VPN, a Raspberry Pi labeled "IT DEPT - DO NOT REMOVE" with its own cellular uplink, a $60 LAN Turtle that captured 17 password hashes on the first business day, and a rogue access point hidden above the ceiling tiles.

None of this shows up on a vulnerability scan. None of it has a CVE. It passes compliance audits every day, because the audit checks whether your controls exist, not whether an attacker can build their own infrastructure inside your walls.

The root cause is a single question most organizations cannot answer: what devices are connected to our network right now, and do all of them belong there?

The tools that turn a sentence into a deployed app are revolutionary for speed, and the same defaults that make them effortless are exactly what leave you exposed. Bolt.new, Replit Agent, Lovable, Cursor, and v0.dev all optimize for getting you from idea to running app as fast as possible. Security pulls the opposite way, so the same holes show up almost every time: row-level security left off, secrets shipped to the browser, missing security headers and input validation, and hallucinated dependencies.

Better prompting closes the easy holes, but verifying is what protects you. We ran the identical audit prompt past two AI assistants on the same naive app: one found five flaws, the other found ten, including a silent fallback that leaked a default customer's record to anyone who mistyped a URL. An automated review is a floor, not a ceiling.

Type a sentence, get a deployed app. The tools work, and people who have never written code are shipping live software with real users and real data. The problem the demo hides is that "it works" and "it is secure" are two different measurements, and the people building this way can only see the first.

The numbers are not improving. Veracode found 45 percent of AI-generated code samples introduced an OWASP Top 10 vulnerability. Georgia Tech counted confirmed CVEs from AI-generated code climbing from 6 in January 2026 to 74 by the end of Q1. And a scan of roughly 380,000 of these apps found more than 5,000 actively leaking sensitive data, because new projects defaulted to public.

One wrong checkbox voided an entire cyber insurance policy. In April 2022, a manufacturer attested on its application that it used multi-factor authentication. A month later, ransomware spread through servers that the MFA never covered. The insurer went to court, and the policy was declared null and void from the day it was signed.

Every "yes" on a security questionnaire is a legal representation, not a guess. In most US jurisdictions, an insurer can rescind the whole policy over a material misrepresentation, whether it was a lie, a guess, or an honest mistake. And it is not only insurance: the same wrong checkbox bites in regulatory safe harbors, vendor questionnaires, customer contracts, M&A diligence, and executive sign-off.

22,468 password hashes.
12,016 recovered.
One afternoon.

After landing Domain Admin on a client network, we dumped every NTLM hash in Active Directory and fed them to our hashcat cluster. Before the day was out, we'd recovered the plaintext credentials for 53.48% of the domain. More than twelve thousand accounts, wide open.

The part that stings: nearly every recovered password was technically compliant. Eight characters minimum. Upper, lower, number, symbol. The audit tool never blinked once while we had the keys to the kingdom.

Here's where the policy fell apart:
-18.5% of recovered passwords hit exactly the 8-character floor, nothing more.
-49.8% had full complexity: letters, a special character, and a digit.
-Only 36% cleared the 12-character bar recommended by PCI DSS 4.0.
-Just 7.6% met the NIST 800-63B Rev 4 single-factor threshold of 15 characters.

Phase 1 used pure mask attacks against the policy requirements: 0.10% recovered. Phase 2 added rockyou with no rules: 1.50% cumulative. Phase 3 layered in the public d3ad0ne ruleset (34,000+ transformations): 13.91%. Phase 4 brought our internal wordlist and custom rules: 50.20% of accounts were unique, for a total of 53.48%.

That proprietary tooling delta tells the whole story. Off-the-shelf public tools left 7,621 passwords standing. Our internal corpus knocked them down in a single phase. No compliance scan, NIST checklist, or audit dashboard can see that gap.

The complete analysis is in the comments: all four phases with full hashcat syntax, length and pattern breakdowns, crack-time projections across five hardware tiers (from a single laptop GPU to our 12-GPU, 4.3 TH/s rig), and framework mapping.

When was the last time someone actually cracked your AD hashes? Not just reviewed the policy. Not just running an audit. Cracked the hashes.