How I work

Authorization comes first

I analyze two kinds of things: systems I own or control, and evidence provided to me under a signed engagement letter that says what's in scope. The first group includes the software my practice runs on, the servers I run it on, and my own computers and accounts.

I don't test, scan or probe client systems, and I don't touch any system I haven't been authorized in writing to examine. If the scope of an engagement is unclear, the work waits until it's clear.

Defense, not exploitation

My work is investigation and loss measurement: reconstructing what happened, measuring what it cost, and recommending the controls that would catch it next time. I don't do penetration testing, and I don't develop exploits.

What specialized AI access is for

Specialized access to AI models, such as access for verified security work, is for two kinds of work:

• Software I own: reviewing the code of the client portal and other software my practice runs on, triaging what my security tests flag before each deploy, and checking that fixes work.
• Evidence provided under an engagement letter: taking apart phishing emails, their headers and suspicious attachments inside the isolated virtual machine described below. Working through exported sign-in logs, audit logs and inbox rules to line the inbox up with the ledger and measure the loss. Writing detection rules for the patterns these schemes leave, such as a new inbox rule, or a change to a vendor's bank details followed by a payment.

Client information stays confidential

As a CPA, I'm required to keep client information confidential. As a tax preparer, I'm also bound by section 7216 of the Internal Revenue Code and the FTC Safeguards Rule, and protected health information from healthcare clients is handled under business associate agreements. In practice:

• In security work, unredacted client records, tax return information and patient information never go to a cloud AI service.
• In that work, I analyze confidential material with AI models on my own hardware.
• When I use a cloud model on part of a case, it sees redacted material only.
• Clients send evidence through a secure channel I set up, not ordinary email.

Evidence is handled so it holds up

I work from copies and keep originals unchanged. I record where each item came from, when I received it and who handled it, so any finding can be traced back to its source if it's ever challenged.

Suspicious files stay isolated

When I examine a suspicious email or attachment, I open it only inside an isolated virtual machine with no connection to client systems or to the servers my practice runs on.

AI access is mine alone

I use AI tools under each provider's usage policies. Any specialized access I'm granted, such as access for verified security work, is for my own defensive work on the two kinds of things above. It stays on my account: I don't share it or resell it, no client or colleague uses it through me, and I don't build it into any product or service, mine or anyone else's.

Problems I find get reported privately

If I find a security flaw in someone else's software, I report it to the vendor privately and give them time to fix it before anything is published. If you find one in mine, here's how to tell me.

Questions about any of this are welcome at andrew@andrewpizzello.com.