How to build your own crypto risk framework (and why most investors don't)?

By November 2021, Axie Infinity looked like proof crypto gaming had arrived. 2.7 million daily players. In the Philippines, "scholarship guilds" rented the game's NFT pets to players who used the in-game token SLP as a wage substitute during COVID. SLP traded around $0.40, AXS (the governance token) hit $165. That priced Sky Mavis above many public game companies. A month earlier, Andreessen Horowitz, Paradigm, and Animoca had led a $152M Series B at $3B.

Four months later, on March 23, 2022, an attacker drained $625M from the Ronin bridge in two transactions. Ronin is the chain Sky Mavis built so Axie could run cheaply outside Ethereum; the bridge is the only road between the two (lock ETH, mint a copy on the other side). The team noticed a week later, when a user complained that their 5,000 ETH withdrawal would not clear.

The hack is remembered as the last nail in Axie's coffin. Forgotten: just before it, SLP had already lost 90%+ of its peak and AXS 75%+, both effectively worthless to late buyers.

Most outlets framed what followed as bad timing. We argue the opposite: the exploit was an accelerant, not the cause. The signals that would have flagged the whole stack sat in plain sight. Web3 just had not standardized them yet.

$625M, Lazarus Group, six days of no monitoring. A format the press knows: large platform, large breach, large loss. Competing-gaming decks ran the same line all year: "Axie was great until the bridge hack."

Pinning failure on one technical event is also convenient for everyone who bought in. If the cause was one piece of infrastructure patched in six weeks, the lesson is narrow: audit bridges, rate-limit withdrawals, expand the validator set. Sky Mavis did all three. Binance led an emergency $150M round in April 2022, with a16z and Paradigm back in on dilutive terms.

On paper, recovery. The company recovered. The thesis the money had bet on did not.

Strip the on-chain layer off and Axie is the most boring grind game online. Breed monsters, fight them, earn for winning. Late-2000s browser-game mechanics, except the token economy held together on one arithmetic condition.

SLP was paid to players for winning battles, and players sold SLP to new players who needed it to breed Axies. The system worked as long as new players arrived faster than the game printed tokens (which made "crypto is a Ponzi" believers happy once the math caught up). By late 2021 the ratio had snapped, and SLP inflated like any overprinted currency.

Bloomberg, reporting Sky Mavis's own numbers on April 4, 2022, showed DAU down 45% from the 2.7M November peak to 1.48M in the week before the breach. SLP had lost 90%+ by February. The Philippine scholarship players the press wrote "future of work" stories about were already drifting away: running scholarship clans in an on-chain grind game had stopped paying.

Two weeks before the drain: wage-substitute use case gone, SLP a 90% loss, AXS off 75%, DAU sliding faster than any patch could slow. The exploit landed on top and finished the job.

The death was demand-side. The headline was supply-side.

On April 14, 2022, the US Treasury attributed the attack to Lazarus Group and added the address holding the stolen funds to the OFAC sanctions list. That turned a hack into a structural problem for everyone downstream.

Once an address sits on that list, every regulated counterparty has to act. Exchanges file suspicious-activity reports on any wallet that touched it. Custodians screen every deposit. Fund administrators verify in writing that LP capital never brushed a contaminated flow. Everyone correlated with Ronin or Sky Mavis paid for it.

The repricing did not stop at AXS holders. Every protocol built on Ronin, every fund whose thesis cited Axie as proof play-to-earn worked, every narratively adjacent project now carried the breach. The press flipped from "future of work" to "is play-to-earn dead" inside a quarter. The whole cohort got repriced as one position, which is what it always had been.

Demand collapse, then exploit, then compliance contagion. Each survivable alone. Stacked, they were not.

The temptation is to call the 2021 record a failure of homework. It was not. The 2021 generation of crypto VCs was not lazy. The industry just had not yet seen enough projects fail to know which categories to check. a16z, Paradigm, Cuban, Ohanian, Animoca, and twenty other investors on Series A and B all ran the standard list: company verification, game assessment, unit economics, team backgrounds.

What none ran was: if the Ronin bridge is drained one Saturday morning, how many other positions in our book do we mark down?

The bridge as a risk category was not yet on anyone's radar. Cross-chain bridges in 2021 were plumbing. Wormhole ($325M) was four months out. Nomad ($190M) was eight. Harmony Horizon ($100M) was seven. The institutional knowledge that bridges concentrate one chain's entire liquidity behind a handful of signing keys, with no withdrawal limits and often no monitoring, was not yet on a checklist. Ronin would be the case that put it there.

Risk frameworks always lag the last incident class by one cycle.

What changes after a clean post-mortem is not the intelligence that it fails this way. Instead, each post-mortem should have produced changes to the checklist that we, as an industry, consider safe. Every named crypto exploit since 2022 has added a row. Wormhole: are bridge contracts upgradable, and who holds the upgrade key? Nomad: does the bridge verify the message it received is the one that was sent? Harmony: how many people sign before money moves, and are any on shared infrastructure? Ronin added all of these and one more: how fast does the team notice a $625M outflow?

That last sits in a different domain from cryptography. A withdrawal-rate-limit (which freezes any outflow above a daily ceiling) would have flagged Ronin in seconds. A 24/7 monitoring rota would have caught it in the hour. It ran six days because nobody was watching.

A usable crypto risk framework is six columns plus the habit of filling them in for every position: Security, Operational, Financial, Dependency, Governance, and Reputation. Each column has yes/no checks tied to mechanisms that would have caught past failures. Ronin sits in the dependency column. The 4-of-9 multisig (four signatures from a list of nine required to move money), with a stale permission granted to Axie DAO in November 2021 and never revoked, sits in the operational column.

A withdrawal-rate-limit check is yes/no. So is "is DAU positive." The question that mattered for Axie in February 2022 was whether DAU was sliding faster than the token economy could absorb without breeding-fee revenue collapsing. That is a judgment call. Two analysts on the same Bloomberg chart could read soft landing vs death spiral. Sky Mavis itself believed in a soft landing.

This is the limitation of every crypto risk framework that relies solely on public data. It can tell you a 45% DAU drop has historically correlated with token-economy failure in 7 of 10 cases. It cannot tell you this one was the eighth. That call belongs to an analyst with a thesis about user behavior, competing games, and what SLP holders are actually doing.

CORE3 names this layer Proof of Voice (PoV): a qualitative review by certified experts that runs in parallel to the quantitative Probability of Loss (PoL) score and never overrides it. The math stays mathematical, the judgment lives on its own track. A failing ecosystem at a specific point in time is the canonical PoV case: observable, defensible, impossible to express as yes/no.

A crypto position carries six categories of risk. On-chain exploit (security). Key exposure from poor operations (operational). Collapse of the project's economic design (financial). Founders with a track record of failing investors (reputation). Exploit through someone else's code or infrastructure (dependency). Ban or restrictive action in core jurisdictions (compliance). Six columns.

The columns are not new. Institutional risk teams have run versions of them on every asset class for forty years. The crypto-specific move is one extra question: which of your positions are quietly running on the same bridge, oracle, or chain. The question Axie's investors did not ask in 2021.

For each column, history gives the failure modes and the controls that would have caught them. Security: was audit scope meaningful, is the code unchanged. Operational: are keys rotated on schedule, are outflows monitored, is anyone on call. Financial: can the token economy survive a 50% drop in new-user inflow. Dependency: which bridges and oracles are load-bearing, and how many positions share them. Reputation: any incident that would price a fund out of an LP relationship. Compliance: is the project structured so a sanctioned counterparty does not force a freeze.

The questions are the product. Each cell has an answer or "I don't know." The "I don't know" boxes are the output: what a position review should ask the team to fill in before sizing.

A framework that asks about positions one by one tells you each vault was built correctly. One that asks about the connections tells you whether the book quietly became one vault. Allocators who run the second are just less correlated than they thought.

Running six columns by hand at allocator scale is not feasible. CORE3 scores 85 metrics across the six domains for 1,463 projects, from load-bearing inputs (validator-set concentration, audit scope, bridge architecture, key rotation, jurisdiction posture) down to boring-but-decisive ones (security-contact disclosure, DNS locking, upgrade-key timelocks). Every input compresses into a single Probability of Loss (PoL) value. Proof of Voice (PoV) sits next to it.

An API release is on the way, so allocators can pull the raw metrics, apply their own weights, and run the framework against their own book. Until then, PoL scores are public at core3.io. Without standardized due diligence and risk transparency, crypto stays in the gray, exposed to preventable incidents and malicious capital extraction.

Break the cycle. Make risk-informed decisions: core3.io