Each study links to a public GitHub artifact. Findings are labelled at the rung actually proven (Observed, Reproduced, Runtime-confirmed, Reported, Remediated, Retested); severities are never inflated to manufacture a result.
A pre-mainnet Soroban options vault trusted a caller-supplied authorizer on the withdraw path, letting an attacker authorize a drain of another account's collateral. Elevated from source-level to Runtime-confirmed with a native cross-contract Soroban PoC (PASS), then reported responsibly. Public finding.
Voting power was read from live balance with no per-proposal snapshot, so one 100-token stake drove 300 votes and defeated a 250 quorum. Confirmed with a cargo-test PoC against the real contracts; snapshot-at-creation remediation posted. Public finding.
solc-js recompile structurally matches the on-chain bytecode; interface enumerated as ERC-20 + Burnable with zero privileged selectors; delivered with a reproducible verify.py and its scope limitation stated plainly. An honest clean verdict backed by evidence, not assertion. Public finding.
Per-key (target, selector) session-key scoping via an owner-signed EIP-712 SessionGrant. Merge-ready src/SessionGrantLib.sol was verified against the real ERC-4337/7579 Execution[] path with forge test 9/9 passing on solc 0.8.33. Public contribution.
On an invited external review of a frozen commit, the in-tree security policy routed reviewers to closed trackers for a superseded review target — a source-binding / identity-drift defect. Proposed a promotion-gate CI assertion so the documented review target always matches the reviewed source. Public finding.
Agent capability scope used an adversarially bypassable substring match and defaulted fail-open. Proposed a typed, segment-exact, fail-closed matcher plus a recomputable AuditEntry so authorization decisions are reproducible. Public finding.