Roadmap

This is OAK's working plan: what is being done now, what comes next, and why. It is updated in the same commit as the work it describes. When an item is done it moves to the log at the end of the "Current work" part, with its commit and the reason it was worth doing, so the file also answers "why does OAK look like this?".

Order of trust: sourced facts first, coverage second. Since the October 2026 source pass, an incident that cannot be traced to a source does not go in, and a mechanism that does not match its source gets fixed before anything new is added.

If you want to contribute against any of these items, open an issue using the relevant template in .github/ISSUE_TEMPLATE/ and the maintainers will comment on scope and ID assignment before you spend effort drafting.

---

Current work

Last reviewed 2026-10-30 (evening).

Next

0. Make incidents replayable end to end: fund flow, wallet links, graphs

Done so far (2026-10-30): the oak-flow data model and check_flow (89365fb); the interactive graph on incident pages (d8cc12f); the address index with infrastructure filtering; the keyless tracer tools/explorer.py (Blockscout, public RPC, Routescan, mempool.space); agent tooling in .claude/ (759272c); and fund flow for Binance, X100 and eight more October incidents (f538c51).

Next small slices. Take one or two per session, then ship and log them:

1. ✅ Trace the rest of October 2026 (14 incidents, 2026-10-30 evening).
2. ✅ Graph polish: label collision avoidance; DEX round trips no longer pull pools left of the receiver.
3. ◐ Binance ↔ Bitget: the attribution is corrected (Elliptic citing ZachXBT, ~$39K; not TRM). Bitget's published receiving addresses are now in its entry. Open: trace both sides to 0x15104a9c… on-chain.
4. ✅ Native hops over RPC: the EOA nonce walk. It found Autheo's Relay bridge deposits and Polymarket's 28.6M POL split.
5. ✅ Solana (public RPC balance deltas) and Tron (TronGrid). Open: XRP, Zcash.
6. Backfill (next session): 61 pre-2026-09 timelines break a rule (check_timeline.py --legacy), and 282 examples have no fund-flow row (--warn). Work in batches by size of loss.
7. Mitigation backfill (what/how/why): about 200 Technique and mitigation sections are generic (check_mitigations.py --legacy). Rewrite them case by case, biggest-loss techniques first. Each bullet needs a concrete control, where it is enforced, and the linked incident it would have stopped.
8. Next rule candidate: a defender quick card. A short block at the top of each incident: the mechanism in one line, the step to check in your own system, the control (linked OAK-Mnn), and addresses to watch (from Key addresses and Fund flow). It should be generated where possible, not hand-written, so it cannot drift from the data below it.
9. Follow-ups found while tracing: follow TranchePool's CCTP burn to its Solana recipient (Solana is now supported); trace the Polymarket POL receivers; trace Bitget's Tron receiver (TBWNgu…, 20.59M TRX in).

1. Write proper entries for real incidents that were filed under the wrong mechanism

Incident	What actually happened (per the correction log)	Quarantined file
Akutars / AkuDreams, 2022-04	Contract bug locked 11,539 ETH ($34M)	2022-02-akutar-nft-influencer-rug
Elephant Money, 2022-04-12	Price manipulation on BNB Chain, ~$11M	2022-04-elephant-money-flash-loan-governance-attack
Fei Rari Fuse, 2022-04-30	Reentrancy, ~$80M	2022-04-rari-fuse-twap-oracle-window-manipulation
Omni bridge on EthereumPoW, 2022-09	Message replay on the ETHW fork (ETHW itself used chain ID 10001)	2022-09-ethereum-pow-fork-replay-cohort
Visor Finance, 2021-12-21	deposit() validation flaw in the vVISR rewards contract minted shares, ~$8.2M	2021-12-visor-finance
Atlantis Loans, 2023-06	Malicious governance proposal on an abandoned protocol spent lingering approvals	2023-06-atlantis-loans
SnowdogDAO, 2021-11	~$30M rug on Avalanche	2023-11-snowdog-dao-avalanche-cross-chain-lock-spoof
Balancer veBAL "Humpy", 2022	Gauge-vote capture	2023-2024-balancer-vebal-delegation-takeover
Fake wallet apps	Kaspersky's App Store "FakeWallet"; fake Ledger Live for macOS	2023-2025-fake-crypto-wallet-google-play-app-store-phishing
Fake security extension	Okta's "ShieldGuard" report	2023-2025-fake-revoke-cash-wallet-security-extension-phishing

Before writing each, run python3 tools/check_known.py and add a ## Root-cause evidence block if the incident date requires one.

2. Clean old mechanisms off technique and actor pages

Known leftovers: OAK-T11.001 (BitMart described with "1Password"), OAK-T17.005 (Inverse), OAK-T5.007 (ICP), OAK-G18 (Karakurt), OAK-G01 (Yapizon / Upbit wording), and anchor descriptions on OAK-T7.007, OAK-T8.004, OAK-T2.003 and OAK-T16.004. Method: for every corrected or quarantined example, grep the corpus for its slug and key claims and align the sibling text.

3. Check suspected duplicates

• EigenLayer airdrop: 2024-05-eigenlayer-restaking-airdrop and 2024-eigenlayer-restaking-airdrop. The first cites no URLs and claims "$1B+" in losses.
• LastPass: 2022-12-lastpass-vault-cohort and 2023-10-lastpass-breach-crypto-drain. The first cites no URLs and ties Chris Larsen's January 2024 theft to LastPass; verify that link and the amount before keeping either.
• Onyx: 2024-09-onyx and 2024-onyx-protocol-year-plus-exposure.

4. Review doubtful technique mappings

• Yearn v2, 2023-02 — maybe OAK-T9.016 (accounting-invariant flaw) rather than its current mapping.
• Compound, 2021-09.
• B2 Network, 2026-07 — OAK-T11.014 may overreach.
• Ankr — the secondary OAK-T9.004.

5. Re-run the external A/B test

6. Coverage gaps below target

7. Waiting on post-mortems

• Pirate Chain (ARRR), 2026-10-14. Filed at T9.014 from the project's statement (issuance-check bypass + mis-parsed output scripts, ~39.8M ARRR). Add the code path when the promised post-mortem is out.
• October 2026 incidents with no mechanism published yet (sweep 2026-10-30). Each is real and sourced for the loss, but nothing yet says how it happened: Affluent (TON, 10-16, ~$1.05M, DefiLlama "unbacked mint" only), Diamante (10-27, operator says breach, no amount or vector), Frogman personal-wallet drain (10-07, ~$4M, vector unknown), MORPHO / EIP-7702 institutional addresses (10-21, ~$2.3M, ZachXBT; key vs. delegation unresolved), HAI malicious proposal (10-06, Optimism, ~$927K at risk, outcome unknown). Also revisit Binance when its post-mortem lands: the "additional points of breach" may move the primary technique.
• Ledger / CryptoBilis final report. The example is filed at T11.007 parent with the implant confirmed; assign a sub-Technique (candidate T11.007.004 in TAXONOMY-GAPS.md) once Ledger ties the losses to the implant.

8. Next content snapshot (~2026-12)

Log

Newest first. Each entry: what, commit, why.

• 2026-10-30 — Mitigations must say what, how and why; Binance cause wording corrected (this commit). Why: the user pointed out that OAK's mitigations were empty words, and that I wrote most of them. They named no incident, and the control that would have stopped Binance's Ethereum takeover was not in OAK at all.
  • New mitigations. OAK-M49: custodial signers refuse EIP-7702 authorizations and watch their EOAs for code. OAK-M50: withdrawal approval sits outside the service that submits withdrawals. Both are tied step by step to Binance, Bitget, Gamdom, Bloom and AlpsCon.
  • Rewritten sections. T13.004 and T15.003 now open with "What would have stopped the 2026 cases".
  • Enforced. tools/check_mitigations.py is in npm run check, and the rule is written into CONTRIBUTING, CLAUDE.md and the skills, so contributors cannot slip generic text through. About 200 legacy sections are listed for backfill.
  • Binance wording. Teng's own post says the attack may have started on a DevOps engineer's laptop, but that it was a multi-step process, not one engineer's permissions, and that no individual is held responsible. "Additional points of breach" appears only in press paraphrases, not in Binance's posts, and is now labelled as such.
• 2026-10-30 17:00 UTC — Binance follow-up (this commit). Why: new statements arrived during the day.
  • Statements and figures. Yi He (04:17 UTC): deposits normal, withdrawals expected within 48 h after "enhancements to the withdrawal service and related address controls". TRM's figure is $6.36B. Ondology saw small bridge tests through MetaMask's router and LI.FI.
  • OAK's own snapshot (block 26306931). Both main ETH holdings (197,460 and 17,260 ETH) have nonce 0, so they have never sent a transaction. The receiver is down to 4.5 ETH after 703 outgoing transactions.
  • Still missing: Binance's full post-mortem.
• 2026-10-30 — All October 2026 incidents traced; tracer covers native hops, Solana and Tron (this commit). Why: the remaining 14 October incidents still said "funds: undisclosed", and the tracer was blind to native coins over RPC and to non-EVM chains.
  • Traced. Seven Sonnet tracer agents covered the 14 incidents, with edges spot-checked here.
  • Corrections found by tracing:
    • NEM ran on 10-21, not 10-22; Slice on 10-20, not 10-21.
    • TranchePool lost ~1.013M USDC once, not twice: the first draw nets to zero. The money left via CCTP to Solana.
    • The Avalanche Safe funds were returned (307,834.86 WAVAX and the rest), confirmed on-chain, against the sources' "unconfirmed". The bot kept 42,541 WAVAX.
    • Autheo's attacker bridged 292.19 ETH via Relay, found by nonce walk.
    • Polymarket's 28.6M POL went to three wallets, now with txs.
  • Attribution fix. The Binance ↔ Bitget address overlap was attributed to TRM via THN. It actually comes from Elliptic, citing ZachXBT, and covers ~$39K. Bitget's receiving addresses were added.
  • Tooling:
    • EOA nonce walk (archive eth_getTransactionCount) for native sends.
    • Solana via public RPC, from balance deltas.
    • Tron via TronGrid.
    • Allowlist additions checked on-chain: Base USDT, wstETH and aBaswstETH; Polygon DAI.
    • The WBNB allowlist entry was wrong, written from memory with a wrong tail. An agent caught it, and it was fixed and verified.
    • Graph labels avoid cards and each other; DEX round trips are laid out as exits.
• 2026-10-30 — Eight more incidents traced on-chain; tracer works where indexers block us (this commit). Why: "where the funds went: undisclosed" was the weakest part of the October set. oak-fund-tracer agents (Sonnet, one per file) traced Initia, zkLink Nova, UNCX, MakerDAO keeper, BiFi, Base vault, Autheo THEO and Polymarket. Their findings were spot-checked against chain data here.
  • Corrections and new facts:
    • BiFi ran on 10-17 05:43, not 10-18, and the MEV bot that took the funds returned 159,103.65 USDC on 10-18. No source mentions the return.
    • Initia's "unmoved" ETH left on 10-27, and 920 ETH went into Tornado.
    • zkLink's first withdrawal came 4 h before the alert.
    • UNCX's "retained" tokens were sold within minutes.
    • Full addresses were recovered: Initia exit, Autheo orchestrator and pair, Base vault hop and bridge.
  • Tracer. Base and Polygon Blockscout sit behind a Cloudflare challenge, which we do not bypass. explorer.py now falls back to public JSON-RPC: it rotates nodes, finds blocks by interpolation, and reads eth_getLogs per allowlisted token, since public nodes require a contract address. Routescan covers Avalanche. Native-coin hops over RPC stay invisible, and the tool says so.
  • Labels. tools/address_labels.yaml now holds canonical tokens, verified on-chain, plus labelled infrastructure. 0x8f10…f996, which shows up in both zkLink and Autheo, is a shared swap executor (~9.96M token transfers), not a cluster link. The address index now skips labelled infrastructure.
  • Site. The full-screen button is highlighted, and the graph uses the full document column.
• 2026-10-30 — The fund-flow graph is an OAK component, not Mermaid (this commit). Why: Mermaid laid 20+ nodes out in an arbitrary order and shrank them to the text column, so the graph could not be read. The new component in src/components/document/fundflow uses the site's design tokens:
  • Columns by stage (funding → attacker → victim wallets → receiver → hops → exits), with neighbour-aligned rows.
  • Role colours for each wallet card, solid lines for transfers traced by OAK and dashed lines for those relayed from a source.
  • Amount and asset labels placed by fan direction.
  • Hover highlights a wallet's links; click opens a card with the full address, a copy button, the explorer link and every transfer with its tx.
  • Zoom, Fit and Full screen, and the view opens on the receiver.
  • The transfer table stays as a collapsible fallback.
  • Checked by screenshot at 1440, 1600 and 390 px.
• 2026-10-30 — Claude Code tooling for OAK agents; X100 traced end to end (this commit). Why: sweeps, tracing and write-ups were done by hand in long sessions, so they came out differently each time, and rules were remembered rather than enforced. The tooling now lives in the repo, so any contributor's Claude session works the same way.
  • Added. CLAUDE.md with the working rules. Skills oak-sweep, oak-write-example (with template.md), oak-trace-funds and oak-validate. Agents oak-source-reader (Haiku, read-only), oak-fund-tracer and oak-example-writer (Sonnet). A PostToolUse hook that validates any edited examples/*.md at once and feeds errors back. A shared .claude/settings.json with allowlists. Offline unit tests in tools/tests that run in npm run check.
  • Pilot. The tracer procedure ran on X100: 15 edges, all checked on-chain. The USDT was swapped for ETH via UniswapX within 6 minutes, the ETH was split to four fresh wallets, and about 2,022.4 ETH entered Tornado Cash in 82 deposits on 10-26 and 10-27. That contradicts the "not moved" in the cited brief.
  • Spillover. The ether.fi entry's Tornado router address is now complete.
  • Note. New agents load only in a new session.
• 2026-10-30 — Fund flow becomes data; OAK traces on-chain itself (this commit). Why: timelines and full addresses were still prose, so nothing could be drawn, queried or joined across incidents, and "where the funds went" stopped wherever the sources stopped.
  • What was added. The ## Fund flow oak-flow block, validated by check_flow.py and exported as oak.json fund_flow. Its edges are marked verified: oak or verified: source. The site draws it with Mermaid and a table of full linked addresses. tools/explorer.py traces keylessly (Blockscout for EVM, mempool.space for BTC, cached). tools/build_address_index.py builds the cross-incident address index (addresses.json), and pages get a "seen in other incidents" block.
  • Binance pilot, traced by OAK. The full initiator address, now confirmed. Three Tornado withdrawals, not three on one day. The 195,460 vs 197,460 ETH discrepancy explained (two transfers). The USDT figure (1.615B) confirmed in the sweep transaction. 14 gas-funded hop wallets. 529 swaps of 35 tokens into 64 DEX pools.
  • Tracer bugs found during the pilot. The tracer first followed counterfeit "ETH"/"USDT" Transfer events that poisoning bots had sent to the attacker's wallets from lookalike addresses (0x8537… imitating 0x8532…), and it would have recorded hops that never happened. Tokens with no price and fewer than 1,000 holders, and zero-value transfers, are now skipped.
  • Still open. The Binance ↔ Bitget address overlap remains TRM's claim, because OAK's Bitget entry has no addresses to check it against.
• 2026-10-30 — Every incident gets an end-to-end timeline with full addresses (this commit). Why: the user could not follow Binance as a sequence, and objected that the lean sweep entries shortened addresses and dropped how the money moved: "attackers have everything and defenders have nothing." The new tools/check_timeline.py in npm run check requires a Timeline table on every single-incident example: entry, preparation, execution, where the funds went and response, with a Technique per step. Undisclosed stages stay in as "undisclosed", and from 2026-09 on no address or hash may be shortened. The rule is written up in CONTRIBUTING, with Binance as the reference (key addresses, a dated funds snapshot, open questions). 47 timelines were added (five Sonnet agents working only from each file's cited sources, reviewed here), and shortened addresses were expanded throughout. Flagged for review: the Flamincome summary differs from SlowMist on the mechanism, and the Qubic restart and fund return lack a fetched source.
• 2026-10-30 — Incidents page filters by month and day (this commit). Why: the site could only filter by year, so "what happened this month" meant scrolling a year of cards. Each example now carries a date: the month comes from the file name, and the day comes from the title only when the title's date falls in that month. 391 examples have a day, 162 a month only, and 130 a year or range. Year → month → day selectors sort by full date and sync to ?year=&month=&day=, so a filtered view can be linked.
• 2026-10-30 — CoinDCX 2025-07 entry vector corrected (this commit). Why: while reading Ondology Labs' other posts for the Binance case. The example said "server breach" and built a teaching point on it, but police and operator statements since 2025-07-31 describe a recruiter lure and malware on an employee's laptop (T15.001 → T15.003). This is the same shape as Binance's DevOps-laptop entry. The rest of Ondology's blog is compliance and tax material, with nothing for OAK.
• 2026-10-30 — Binance: EIP-7702 leg verified on-chain (this commit). Why: the T13.004 secondary rested on TRM as relayed by The Hacker News. Ondology Labs recovered the signer of all 50 authorizations as the Binance 14 EOA and published tx hashes, the rehearsal on test wallets, the Hot Wallet 34 payout of 823M USDC and a per-chain breakdown. The mechanism is now corroborated, and the remaining unknown is narrower: how the service container got Binance's keys to sign.
• 2026-10-30 — Month-end sweep 2026-10-17 → 10-30: 23 incidents (this commit). Why: October is the worst month on record and the corpus stopped at 10-17. SlowMist tracker first, then RektRadar daily briefs, which surfaced five earlier-October gaps the tracker's first page did not show (THEO, Kolan, BiFi, AlpsCon, MetaMask Staking). Headline: Binance ~$6.3B (T15.003 from Binance's preliminary finding; T13.004 from TRM's reconstruction of 50 pre-signed 7702 authorizations; G01 attribution is the analytics firms', not Binance's). Polymarket's second frontend-vendor drain in four months (DebugBear RUM script, session hijack rather than approval prompts). Initia is the second "VM bug paid at par by the bridges" after Liquid. Six new T9.016 anchors. Five held under Next §7 for want of a mechanism.
• 2026-10-17 — Sweep 2026-10-11 → 10-17: four incidents, Ledger implant confirmed (this commit). Why: keep recency tight. Demex (2026-09-25, missed last sweep; T10.002 from the operator's post-mortem), UNCX VestingManager (T9.016), Gamdom (T15.003 + T11.011, the second casino multi-chain hot-wallet drain in three weeks), Coldcard's X account (T15.006). Ledger's confirmation of a hardware implant moved the CryptoBilis entry from disputed to corroborated and opened the T11.007.004 candidate. Pirate Chain held until it has a source.
• 2026-10-17 — Sweep follow-up: five more from the SlowMist tracker (this commit). Why: the first pass used news and vendor posts only and missed the tracker, as the sweep-coverage rule warns. Pirate Chain (T9.014, now sourced), Papertrade (T17.001), X7 Finance (T9.016), Set Protocol (T9.011), Based Visa Card KYC exposure (T15.004); UNCX loss raised to ~$1.49M. Rule for next sweeps: read hacked.slowmist.io first.
• 2026-10-11 — OAK-T9.012 deprecated (d65a2b3). Why: the mechanism ("buy before the pool has liquidity") cannot happen, and its three anchors were invented. The real neighbouring patterns, launch sniping and pool pre-creation at a skewed ratio, went to TAXONOMY-GAPS.md as candidates.
• 2026-10-11 — Every example now cites a source, and the build enforces it (1c98e22, with b11ee35, d5b4bb5, a64a876). Why: the KelpDAO and THORChain cases showed that OAK held confident entries nobody had checked. Unsourced examples went from 253 to 0. 78 files with no record of the incident were moved to quarantine/, duplicate sets were merged, and ~150 were corrected in place. check_sourcing --ci --max 0 now fails the build on any new unsourced example. Examples: 738 → 651.
• 2026-10-11 — root_cause_evidence, withdrawn examples, oak-mcp 0.4.0 (5120e04, d4b34c9). Why: an external A/B test showed a model using OAK could not tell how solid a root cause was, or that an example had been withdrawn. Examples dated 2026-10 onward must state technique, source, source type and confidence. oak.json lists withdrawn files with the reason, and oak-mcp exposes both.
• 2026-10-11 — THORChain reclassified, OAK-T10.008 deprecated (591e636). Why: THORChain's own Exploit Report #2 attributes the May 2026 loss to key extraction through its threshold-signing library, not the observer-signature flaw OAK had recorded from a same-day third-party analysis.
• 2026-10-11 — New techniques OAK-T9.016 and OAK-T11.014 (15d2c9b). Why: contract accounting-invariant flaws and operator privileged-key compromise had been spread across techniques that did not fit them, which blurred both statistics and mitigations.

---

Longer-range plan

Written at v0.1. Reviewed 2026-10-12: delivered items are struck through. The website and governance items were not re-checked.

v0.x — content depth

• WalletConnect session hijacking — delivered as OAK-T4.006.
• QR-code phishing as an entry vector — no standalone Technique yet; add one only if a sourced incident shows a QR-specific mechanism.
• Privacy-chain hops (OAK-T7.005) and DeFi yield-strategy laundering (OAK-T7.006) — delivered.
• Standalone T6 entries — delivered: fake audit claim (OAK-T6.002), audit-pending marketing claim (OAK-T6.004), renounced-but-not-really (OAK-T1.003).
• Per-Technique calibration / scope refinements — not re-checked:
  • OAK-T1.001: governance-gated sub-classification for governance-timelocked modifiable fees.
  • OAK-T1.002: per-extension sub-Techniques for other Token-2022 extensions.
  • OAK-T2.002: cross-chain locked-liquidity setups.
  • OAK-T3.001: reference funder-graph depth and thresholds per chain.
  • OAK-T3.002: per-venue cycle-window and net-flow thresholds.
  • OAK-T3.003: confidence-band reporting convention.
  • OAK-T5.002: calibrate window size and cumulative-fraction threshold; move reference implementation from gap.
  • OAK-T5.003: chain-specific sub-Techniques if detection diverges.
  • OAK-T8.001: cluster freshness/decay convention.
• Per-Technique Data Source backfill — technique pages do not yet carry explicit Data Sources references.

---

v0.5 — multi-implementation matrix and co-maintainers

• Two-dimensional COVERAGE.md (Technique × Implementation) with at least three implementations represented (operator-behaviour detectors + protocol-layer monitors + audit-firm continuous-monitoring services).
• Cross-walk between OAK Techniques and OWASP Smart Contract Top 10 categories — delivered as CROSSWALK.md.
• Invite 2–3 co-maintainers from the contributor community to share PR review load, per the v0.1 commitment in CONTRIBUTING.md.
• First retroactive coverage audits — comparing self-attested vendor coverage entries against publicly verifiable evidence.
• Move the per-Technique example bar to "≥1 example per Technique." — met: every active Technique has an anchor (BACKLOG.md).
• NFT-specific Techniques — delivered as Tactic T12.
• Detection-test corpus — labelled-example harness derived from the cited academic datasets.

---

v1.0 — stable taxonomy

The hard semantic commitment that comes with v1.0:

• The Tactics are stable. New top-level Tactics may be introduced in v2.0+ but not before, and only after community consultation.
• Existing Technique and Group IDs are stable. Deprecated entries are marked Maturity: deprecated and remain in the repository with a pointer to the superseding entry. IDs are never reused.
• A canonical machine-readable schema is published alongside the markdown, frozen as stable. tools/oak.json was promoted from a draft to a released (pre-1.0) schema at v0.7; v1.0 freezes it as the stable, ID-stable contract.
• Account abstraction (ERC-4337) coverage — delivered as Tactic T13.
• Validator / staking / restaking coverage — delivered as Tactic T14.
• Severity and timing dimensions on every Technique.

---

Website MVP backlog (post-Phase R)

The Vite + React static-site MVP at onchainattack.org exposes the corpus and was extended in Phase R to render Mitigations and Software workspace views. Remaining work to make the site a first-class consumer of the OAK relationship graph:

• Per-Technique detail page — render related Mitigations (incoming mitigates), Software (incoming uses), Worked Examples, Citations, and Data Sources as a single technique-centric panel; replace the current document-viewer-only path.
• Per-Mitigation detail page — list every Technique the Mitigation applies to as a clickable matrix slice; surface Audience and Class as filterable facets.
• Per-Software detail page — render Used by Groups, Observed Techniques, and lineage links (predecessors / successors) as graph-style cross-references.
• Per-Group detail page — surface Software the Group uses (G→S edges) and the Techniques those tools cover.
• Filters by relationships — facet the Matrix view by "techniques mitigated by M-NN" / "techniques used by S-NN" / "techniques observed in G-NN" so a risk team can build coverage views without manual cross-referencing.
• Search index expansion — current search covers Techniques, Examples, Actors, and Docs; extend to Mitigations and Software entries plus full-text across ## Description content.
• Relationship graph visualisation — a force-directed or layered visualisation of mitigates and uses edges so the cross-axis structure is browsable, not just queryable.
• STIX 2.1 export — delivered. tools/export_stix.py emits tools/oak-stix.json (601 STIX objects: tactic + attack-pattern + course-of-action + malware/tool + intrusion-set + data-source + relationship SROs). UUIDs derived deterministically via uuid5 from a pinned OAK namespace so STIX IDs stay stable across regenerations.
• Per-Tactic kill-chain banner — the matrix currently lists Tactics as columns; add a phase-ordered banner that shows the canonical T1→T8 kill-chain progression and the auxiliary T9-T14 tracks visually distinct from it.
• Coverage drill-down — current Coverage view links to COVERAGE.md only; add an interactive matrix showing per-Technique status (full/partial/gap) with vendor-implementation row breakouts as the reference-implementation count grows in v0.5.
• Bundle size — Vite warns the production JS bundle is over 500 kB minified (~617 kB gzip). Code-split by view (React.lazy per workspace section) to land under 250 kB initial.
• Code splitting for src/data/generated.ts — the ~7 k-line generated TypeScript constant currently inlines all document HTML; move large markdown bodies to per-document JSON fetched lazily on document-viewer entry.

These items do not gate the v0.1 content cut. They are the work to land alongside or after the v0.1 launch announcement.

---

Repository-level commitments (out of release scope)

These items track strategy-level work that progresses independently of content versioning.

• Coverage-claim verification methodology — community-driven RFC on how vendors and detection tools can self-attest to OAK-Technique coverage in a comparable, machine-readable way. The detail is open for community input; the goal is honest comparability across implementations, not certification gatekeeping.
• Governance research — interviews with comparable open-framework communities (CIS Controls, OWASP Foundation, Linux Foundation) on long-term sustainability and incorporation patterns. Outcome posted as a community discussion document.
• Trademark filings — defensive word-mark registration for the OAK name post v0.1 launch, to prevent third-party squatting.
• Visual identity — SVG wordmark/logo for the README banner and onchainattack.org landing page.

---

How to propose changes to this roadmap

This file is owned by the maintainers and is updated alongside the per-Technique Discussion sections from which it is aggregated. To propose a new v0.x work item, open an issue with the appropriate .github/ISSUE_TEMPLATE/ template; if the proposal is accepted, the maintainers will reflect it here at the next roadmap update.