No victim in this one, and no decision to second-guess. It is here because the shape of the number is the lesson.
Chainalysis estimated a record $17B stolen through crypto scams and fraud in 2025, having directly attributed at least $14B of it on-chain. Impersonation schemes grew more than 1400 percent year on year. Their average payment severity rose more than 600 percent. Across all categories the average scam payment went from $782 to $2,764.
Operations using AI earned 4.5 times what the others did, $3.2M against $719,000, across about nine times the daily transaction activity.
Now hold that against what security advice is about. Hardware wallets, firmware, key derivation, contract audits, cold storage.
The decision point is structural rather than personal. The category that takes the most money is conversation. Conversation is the category the advice spends the least time on.
Most security advice in crypto arrives as a checklist. Use a hardware wallet. Do not click links. Never share your seed. The advice is not wrong. It is floating in space, disconnected from any picture of who is coming for your coins and how.
Without that picture the checklist feels like superstition, and superstition gets abandoned the first time it is inconvenient. With it, every item snaps into place and you can tell which ones you actually need.
That picture has a name and the name sounds more technical than the thing is.
A threat model is one sentence
These are the people who would take my coins, this is what they would have to pull off, and this is what I do about it.
Adversary. Requirement. Countermeasure. Once you have written that honestly for your own situation, every security decision reduces to a single question: does this raise the cost for the adversary I actually have, or am I doing busy work?
The most common beginner mistake is borrowing someone else's. The model that fits a state-targeted journalist is not the model that fits a parent with $40,000 in savings. The first will cost years of your life and all of your patience. The second will get the journalist hurt. Pick yours.
The fifteen patterns
Everything you will meet fits in three families. The families matter more than the individual entries, because the family tells you which control applies.
| # | Pattern | Family | What it needs from you | The control that stops it |
|---|---|---|---|---|
| 1 | Fake support, inbound | Conversation | You accept help you did not seek | Source support yourself, from the app or site you already had |
| 2 | Impersonated authority | Conversation | You act on a voice, face, or title | Out-of-band verification on a channel you opened |
| 3 | Romance and long con | Conversation | Weeks of rapport before the ask | Notice the destination, not the pace |
| 4 | Send first to receive more | Conversation | One outbound payment | Refuse the pattern absolutely, no exceptions |
| 5 | Fake platform with working small withdrawals | Conversation | Escalating deposits | Treat any return you cannot verify on-chain as fiction |
| 6 | Recovery services after a loss | Conversation | A fee, paid by an existing victim | Nobody who contacts you first can recover funds |
| 7 | Fake job, interview, or paid task | Conversation | You run their code or file | Never execute anything from a hiring process on a funded machine |
| 8 | Malicious token approval | Authorisation | One signature granting spend rights | Read the approval, cap it, revoke on a schedule |
| 9 | Blind signing an unreadable payload | Authorisation | Approval of a hash you cannot decode | Refuse to sign what the device cannot render |
| 10 | Address poisoning and lookalikes | Authorisation | A copy-paste from history | Verify the full address, never the truncation |
| 11 | Drainer front-end or paid search ad | Authorisation | Connecting and signing on a real-looking site | Reach every dApp through your own bookmark |
| 12 | Seed exposed on your own surface | Compromise | A photo, a cloud note, a password manager | The seed never becomes a file, ever |
| 13 | Account takeover by SIM swap or email | Compromise | Control of your phone number or inbox | Carrier lock, hardware keys, and no SMS as a factor |
| 14 | Malware, infostealers, malicious dependencies | Compromise | Execution on the device holding keys | Device separation, and keys on hardware |
| 15 | Physical coercion | Compromise | Knowledge that you hold | Disclosure discipline, and a plausible smaller wallet |
Patterns 1 to 7 are conversations. Patterns 8 to 11 are signatures. Patterns 12 to 15 are your devices and your identity.
Read down the last column and notice something. Almost none of those controls are products. They are procedures. This course teaches the procedures and F105 already taught you four of them.
The ordering is the point
For an individual holder, roughly in the order these actually happen:
- Conversation, by a wide margin. Someone convinces you to type, paste, or sign. This family dwarfs the others combined, and the two largest individual thefts ever recorded, $282M in January 2026 and $243M in August 2024, were both support impersonations.
- Seed exposure to your own digital surface. A photo in the camera roll, a cloud-synced note, a password manager vault. Self-inflicted, extremely common, and F106-05 shows what happens when a vault provider is breached.
- Authorisation errors. You sign something that does not say what you thought, or you paste an address from a poisoned history. F105 is the whole course on this.
- Lost access with no backup. No attacker at all. A dead device, a forgotten passphrase, one piece of paper. The chain does not care that nobody meant harm.
- Account takeover. SIM swap or email compromise, which matters most if you keep value at a custodian.
- Malware and supply chain. Real, and mostly consequential when keys live on a general-purpose machine.
- Physical coercion. Rare, rising, and it changes your entire model if it is genuinely on your list.
Run this exercise honestly: for each item, how much of your current behaviour actually addresses it? The usual answer is that the top three are what reaches real people, and the bottom three get most of the attention because they are more interesting to read about.
Reports state that the average scam payment rose 253 percent, from $782 to $2,764. Percentage changes are the single most commonly mangled figure in this industry, so verify it rather than repeating it.
The multiple:
$2,764 / $782 = 3.53 times
The percentage increase, which is the multiple minus the original:
(2,764 - 782) / 782 = 1,982 / 782 = 2.534 = 253.4 percent
Both describe the same fact. "Up 253 percent" and "3.5 times larger" are consistent. "Up 353 percent" would not be, and neither would "3.5 times more than it was", which is a different claim by one whole multiple.
This matters beyond pedantry, because the same slip is what makes a marketed return look twice as good as it is. Whenever a percentage is doing persuasive work in a sentence, convert it to a multiple and check that both readings agree. If the source cannot survive that, stop reading the source.
Two threat models, side by side
Two people, both holding, with genuinely different correct answers.
A. Thirty-five, an engineer, $25,000 in a wallet. Low public profile. A few crypto follows, nothing that announces holdings. Stable jurisdiction.
Realistic threats: conversation, self-inflicted seed exposure, address poisoning, lost access. A reasonable setup is a hardware wallet, a metal backup in a known location, nothing in the cloud, and a habit of verifying every signature on the device screen. They do not need to plan for coercion. They should assume someone impersonates their wallet's support team this year.
B. Fifty, founder of a public crypto company, several million across wallets. High profile, named team, publicly known addresses, travels often.
Everything A faces, plus targeted phishing built specifically for them, plus physical risk. The setup needs multi-signature, geographic key distribution, household awareness, a smaller wallet that can plausibly be the whole story, and an explicit plan for the worst day. They do need to plan for coercion, and cannot pretend otherwise without lying to themselves.
B's setup applied to A is expensive overhead protecting against nothing. A's setup applied to B is negligence. Same chain, same assets, opposite answers.
You are not A or B. You are somewhere on a line between them, and the point moves. Holdings grow. A profile becomes public. A job changes. A local news story names you.
Your threat model is not a one-time exercise, and the most dangerous moment is the one just after you cross a line you did not notice.
The two questions
Before adding a step, buying a device, changing a backup, or following a guide:
Which threat does this actually reduce, for me? If you cannot name it in one sentence, it is probably theatre.
What new threat does this create? Every safeguard adds surface. A deposit box adds a third party. Multi-signature adds a coordination problem. A cleverly hidden device is something you can forget, and F110 is full of people who forgot. Make the trade consciously.
Applied honestly, those two questions keep most people from over-engineering and under-protecting at the same time, which is the usual condition.
The serious threat is a hacker breaking my wallet's cryptography.
This is where the attention goes and it is close to the bottom of the list. Nobody is factoring your private key. The cryptography is the part of this system that works.
What actually happens is that you are asked, convincingly, to hand over the key or to authorise a transfer yourself, and then the cryptography works perfectly on the attacker's behalf. The strength of the encryption is not in question in any incident in this course.
The uncomfortable implication is that the defended surface is you, and you cannot be patched. You can only be given procedures that do not depend on your judgment holding up under pressure. That is what the rest of this course builds.
What this course does
F106-02 covers why capable people fall for the conversational family, because knowing the list does not make you immune and pretending otherwise is how learners get hurt. F106-03 and F106-04 cover the surfaces you control. F106-05 covers how attackers actually build target lists, which is less personal and more automated than most people assume. F106-06 is what you do in the first hour after something goes wrong.
Then the gate, F106-L, The Gauntlet: six to ten simulated attacks across email, in-app notification, a sandboxed dApp, a mock DM and at least one synthetic-media channel, delivered over fourteen days without warning. Falling for one triggers a teardown and adds attempts. The pass standard is a 100 percent catch rate on critical-severity attempts, because in this domain a 90 percent catch rate is a total loss with extra steps.
A threat model is one sentence: who would take my coins, what they would have to pull off, and what I do about it. Every attack fits three families, conversation, authorisation and compromise, and the conversational family takes more money than the other two together, which is the exact inverse of where security advice points. Rank the threats you actually face, ask what each new measure protects and what it costs, and never borrow the model of someone whose life does not look like yours.