TLDR: On October 13, a fake “Please pay via ACH today” email signed with my co-founder’s name reached Symbol’s accounts payable inbox. A spam filter held it, so nobody who pays invoices ever saw it. The filter learned from the attack. The people did not. In the same two weeks, Sublime Security shipped simulations built from the attacks it detects in each customer’s mail, and KnowBe4 added custom deepfake scripts that can use internal project names and phishing tests personalized from coaching data. The market is converging on one idea: the best drill is the attack that already reached you. MSSPs can run that loop today: harvest what got caught, strip the payload, keep the pattern, aim it at the role it targeted, and measure reports, not just clicks.
Tuesday, October 13, 4:51 PM in Bogotá. Google Groups sends me a moderator’s spam report for [email protected]. One message is waiting in the queue.
The subject line is “RE: Vision Is Cheap. Translation Is Everything.” That is not a thread we ever had. It reads like a conference talk title, which is exactly why someone in finance might assume they missed the original. Under it, three lines:
“Please pay via ACH today.
Thanks Craig Sandman”
Craig is my co-founder. He does not ask anyone to pay anything via ACH today. Below the signature sits a quoted “earlier message” from a consultant at a domain we have never worked with, dated the previous Friday, addressed to Craig, to make the request look like the tail end of a conversation that had been going on for days.
The filter caught it. If nobody approved it, it would be discarded automatically. Case closed, by any normal standard.
Except that the people the email was built to fool, the ones who open accountspayable@ and actually move money, never saw it. They will not recognize the next one, because as far as their experience goes, this one never happened.
The Filter Learned. The People Did Not.
Every security stack has this quiet asymmetry. When an attack is blocked, the system gets better: a sender reputation drops, a rule fires, a model adds an example. The humans downstream get nothing. No story, no pattern, no “this is what it looks like when someone uses Craig’s name to rush a payment.”
That is fine until the filter misses. And filters miss most often on exactly this kind of message. There is no link, no attachment, and no malware. It is a short, polite instruction from a name you trust, inside what looks like an existing thread. Business email compromise does not need to beat the filter every time. It needs to beat it once, on a day when someone in finance is tired and the request sounds routine.
So the question is not whether our filter did its job. It did. The question is why the most realistic, best-targeted phishing lesson our finance inbox received all month went straight to a moderation queue and stayed there.
Two Weeks, Three Vendor Moves, One Idea
The awareness market reached the same question in public this month.
- October 6: Sublime Security made Tailored Phishing Simulations and Training generally available. It ranks simulation templates by the attack techniques most active in each organization’s mail flow, and lets teams use AI to turn a malicious email it detected into a simulation “in minutes.” The pitch, in its own words: “the best phishing tests reflect the attacks your employees are most likely to see next."[1]
- October 13: KnowBe4 added custom scripts to its AIDA deepfake videos, up to 1,000 characters, and told admins they “can use internal project names and team terms to make your training more realistic."[2]
- October 14: KnowBe4 also started using real-time coaching event data from connected security tools to personalize the phishing tests AIDA creates for each user.[2]
Three releases, two companies, the same direction: stop testing people with what an attacker might send someday, and start testing them with what an attacker already sent, or what your own environment says they will send next.
I think that is right. I also think most MSSPs can do the useful part of it this month, by hand, without a new console.
Our Own Drill Got Treated Like a Real Phish
The other scene that convinced me is less dramatic and more useful.
Wednesday, October 7, 12:26 PM. Andres, at Wawandco, a software company I also run that uses Symbol’s simulations, gets an email from “Google Slides Support” at [email protected]. “Jhon Smith has invited you to edit the following document: Summary Slides.” Zeros where the o’s should be, and Google’s Mountain View address in the footer.
It was a Symbol simulation. Andres did not know that. At 1:15 PM he forwarded it to me with two words: “Looks sketchy.” I passed it to our report address six minutes later, and Andres sent it to the team’s phishing inbox a few minutes after that.
That is the outcome we want, and it is still a template outcome. A fake Slides invite is a pattern every attacker uses, and it is good practice. But it is generic. Nobody built it for Andres. The ACH email, on the other hand, was built for us: a real co-founder’s name, a real role inbox, a payment instruction, and a fabricated thread. If I could only send one of those two emails to a finance team next week, I know which one would teach more.
Why the Real Phish Makes the Better Drill
Generic templates are not useless. They build the basic habit: pause, check the sender, report. But real attacks carry four things a library cannot.
The attacker already did the research. Someone looked up who runs Symbol, guessed which inbox handles payments, and decided a recycled-thread format was worth the effort. Your drill inherits that targeting for free.
The pattern is current. A library template reflects what was common when it was written. The message in last Tuesday’s quarantine reflects what is being tried against your client right now.
The target is obvious. The ACH request was aimed at accounts payable, not the whole company. That tells you who should get the drill. Sending a wire-fraud lure to the engineering team teaches them something. Sending it to the two people who can actually release a payment teaches the people who matter.
The story is true. “This exact email reached our finance inbox last Tuesday” lands differently from “here is an example of business email compromise.” People pay attention to things that happened to them.
The Loop an MSSP Can Run This Month
You do not need detection data piped into a simulation engine to start. You need a habit that connects the quarantine to the training calendar. Here is the loop I would run.
1. Harvest what got caught
Once a week, look at what your filters held for each client: quarantine, moderation queues, user reports, the messages your SOC already triaged. You are not hunting for volume. You are hunting for the one or two messages that were clearly built for that organization: a real executive’s name, a real vendor, a real role inbox.
2. Strip the payload, keep the pattern
Never resend the attack itself. Remove the links, attachments, payment details, and any third-party names that might belong to real people or compromised businesses. Keep what made it convincing: the role targeted, the sender it impersonated, the thread format, the urgency, the ask. “Pay via ACH today, signed as the CEO, in a fake reply chain” is the pattern. The attacker’s domain is not.
3. Aim it at the role, not the roster
Send the drill to the people the attack was aimed at, plus the people who back them up when they are out. For a payment request, that is accounts payable, the controller, and whoever can approve an exception. Small, targeted, and soon after the real attempt.
4. Agree on ground rules before you get realistic
KnowBe4’s own release notes include a content check before a custom deepfake script can generate audio.[2] That instinct is worth copying even without the tool. Decide with each client which executives can be impersonated, which internal project names are fair game, and which topics are off-limits, such as layoffs, bonuses, or health benefits. A drill that feels like a betrayal teaches people to distrust the security team, not the attacker.
5. Measure the report, not just the click
The ACH email has no link, so there is nothing to click. Click rate cannot score it. What you want to know is whether finance reported the request, how fast, and whether anyone tried to verify it through a known channel, like calling Craig instead of replying. Report rate and time to report are the metrics that fit the attacks people actually get.
6. Close the loop with the real story
After the drill, tell the group the truth: this was based on a real email that reached your inbox on a specific date, and here is what gave it away. That one sentence turns a test into a briefing.
Where Symbol Fits
I want to be careful here, because “turn real attacks into drills” is a feature claim, and we have not shipped Sublime’s version of it. What Symbol does today covers most of the loop above.
Our phishing simulation team watches for active campaigns and builds new simulations within days, so library templates track what attackers are using now. When a client needs a specific scenario, like a payment request signed as a named executive, MSSPs can request a private custom template that only their organization or partnership can see, with custom variables and placeholders. When someone clicks, training matches the simulation they fell for, so a wire-fraud lure leads to wire-fraud training. And Report a Phish gives every user a one-action report path, a button in Outlook or a forward address for Gmail, with simulated reports tagged automatically so analysts are not triaging their own drills.
The part no product can do for you is the weekly look at the quarantine and the judgment about which caught attack deserves to become next week’s lesson. That judgment is the job.
For the record, Symbol’s training platform reminded me this morning that I still have five modules pending, including “Boeing Wire Fraud” and “CSG HR Is Calling.” The co-founder whose name the attackers borrowed for a wire request would probably like me to start with the first one.
The Attack Was Free Training. Use It.
Attackers spent real effort on that ACH email. They picked the inbox, borrowed a name, invented a thread, and kept it short enough to feel routine. All of that work landed in a moderation queue and was about to be deleted automatically.
The vendors shipping this month are right that the most relevant simulation is the one built from what actually reached you. You do not have to wait for that pipeline to arrive in your stack. Open the quarantine, pick the message that was built for your client, take out the payload, and send the pattern to the people it was meant for.
The filter already learned from that email. Your clients’ people should too.
References
- Sublime Security, “Sublime Security Launches Tailored Phishing Simulations and Training,” PR Newswire, October 6, 2026. https://www.prnewswire.com/news-releases/sublime-security-launches-tailored-phishing-simulations-and-training-302899306.html
- KnowBe4, “Artificial Intelligence Defense Agents (AIDA) Change Log,” KnowBe4 Knowledge Base, entries for October 13 and October 14, 2026. https://support.knowbe4.com/hc/en-us/articles/38964045370899-Artificial-Intelligence-Defense-Agents-AIDA-Change-Log
