dragonked2 / known-incidents

KNOWN INCIDENTS (reasoning examples, not checklists)

Each entry: root cause → class → primitive → violated invariant →
prerequisites → sequence → asset-flow pattern → why naive scanners miss
it → detection strategy → analogues. Do not invent incidents; verify
against reliable sources before citing.

Oracle manipulation

• Mango Markets ($117M): thin-perp price pushed with attacker capital,
  inflated collateral → undercollateralized borrow. Invariant:
  debt <= collateral x LTV evaluated on attacker-set price. Detect:
  price influence → state influence → asset influence with viability math.

• Curve ($70M class): read-only reentrancy + stale price view feeding
  valuations. Scanners miss it: no classic reentrancy write pattern.

Accounting / shares

• ERC4626 donation/inflation: first-depositor or donated
  totalAssets skews shares = assets x supply / total. Detect: empty
  vault + raw-transfer-receivable + missing virtual offset.

• Desync (paired state): one path updates A without B (early return,
  fast path). Detect: writer enumeration per pair.

Bridge / signers

• Wormhole ($320M class): unverified/uninitialized upgrade path
  minting unbacked assets. Invariant: minted <= verified locked.
  Detect: mint-authority enumeration, not just message parsing.

• Multisig/quorum failures: duplicate-signer counting, threshold
  confusion, rotation gaps. Detect: distinct-signer tests.

Access / proxy / signature

• Parity freeze ($150M+ class): unprotected init/ownership seizure.
  Detect: initializer coverage on every ownership-bearing contract,
  including implementations.

• Permit replay: missing nonce/chainid/address binding reused across
  contexts. Detect: digest-field checklist per signed action.

How to use

Analogize structure, not names: map the incident's asset-flow pattern
and violated invariant onto the target's model. If the invariant and
the attacker-controlled assumption both transfer, you have a hypothesis
worth testing — not a finding until gated.