Marius CiepluchandClaude Opus 5
c0427e6ceb
NodeGoat's documented insecure direct object reference reads with allocationsCol.find(searchCriteria()), and `find` was absent from an operation table holding findOne, findById, findAll and findByPk. The obligation was missing at the commonest read there is, so that defect raised nothing at any point in the pipeline. It was absent for a good reason, and the reason is the argument, not the name. Measured on wp-statistics, a WordPress plugin that ships jQuery: its assets hold 88 calls named find, of which 70 pass a jQuery selector, 13 pass a predicate to Array.prototype.find and 5 pass a selector built at run time by a template literal or a concatenation. None is a collection read. So the obligation is claimed only where the first argument has the shape of a query document. The operation fact states which argument with query_document_arg, and the arbiter that can see argument shapes decides it: no arguments at all, an object literal, or a builder call are claimed; a method reference, a literal, an operator call and a bare identifier are not. Each exclusion is one of the measured classes above rather than a guess. Zero of those 88 calls became an obligated operation, and the zero is measured rather than empty: the same run's census reports 60 files and 4,552 methods, and the files holding those calls are in it. On NodeGoat all three find calls are collection reads and all three are claimed, which raises the obligation at four more handlers -- the allocations handler that takes userId from req.params on an authenticated route, the benefits pages whose isAdmin guard is commented out in the route table, and the memo listing. Seven of fifteen handlers now reach a verdict, all corroborated, none reported without a judge, and the census is unchanged on both subjects. A bare identifier is deliberately not claimed: col.find(query) and arr.find(pred) are the same shape and neither repository measured holds one, so this is a stated false negative rather than an unmeasured false-positive risk. 1,415 tests pass, sixteen skipped; twelve Joern-gated query tests pass in the image. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]> Claude-Session: https://claude.ai/code/session_01EmM1j46qgKMh3zqD1QMwEo
13 KiB
Pre-push safety net status — v1.2.0-alpha.1
Updated 2026-09-16. Phase: implementation. Requirements, design and tasks are approved. 28/29 executable tasks complete; task 8.3 remains blocked. Rollout: NO-GO. The alpha tag publishes experimental implementation and evidence; it does not approve normal hook capabilities or production enforcement.
Next release milestone
M1a is met and the M1b comparison now works through the production rules. The other four milestones remain unverified, and rollout remains NO-GO.
Every veto M1a recorded has been repaired. On 2026-09-16 the production comparison classified
the pinned NodeGoat revisions without any lifting: the vulnerable revision is new with the
three eval witnesses attributed to the change that introduced them, the fixed twin is
analyzed and quiet, and the unchanged revision is unchanged. Production and the recorded view
agree exactly, which is how we know the vetoes were repaired rather than bypassed. No capability
is enabled, no defect is admitted and no alert is emitted. See
M1a and M1b evidence.
The exit run is recorded. Every veto M1a listed is repaired, plus two the repairs exposed: the
comparison's own degradation and boundary checks, and the runner handing every candidate the
whole transaction's gaps. On the vulnerable revision admission now rejects on
capability_unavailable alone.
One half of M1b remains. The milestone also asks for a developer-readable finding carrying a supported security consequence and a concrete repair direction. Both are owned by an evaluated capability declaration and the registry is deliberately empty, so nothing renders them. Writing that text without a declaration is what admission exists to prevent, so it is not worked around. Task 10.10 renders it through an explicitly experimental declaration with eligibility still disabled.
M1a ran on 2026-09-16 within its one-week box. The raw <lambda>2 witnesses are present at
contributions.js lines 32–34, the fixed twin was genuinely analyzed and quiet, and the
unchanged revision was not reported as a new regression, so the kill criterion did not
trigger. Five vetoes fired and were recorded; with all five lifted the detection still stops
at operation_correspondence_unresolved, because a newly introduced operation sits on a
changed line and has no base anchor. That fifth veto was not in the earlier diagnosis and is
now M1b's central repair. See M1a evidence.
The 28/29 task count is implementation progress, not proximity to rollout. Follow the accepted release milestone spec: usefulness → speed → shared PHP/Python transfer → independent qualification → narrow advisory rollout.
Security capability, 2026-09-17
Group 12 closed three defects that limited what the tool can establish, each measured on the pinned NodeGoat revisions with detections unchanged throughout.
- Access control could not answer anywhere. Dominance and configuration were never asked for
change context, because the evidence pass dispatches taint only. Each family now describes its
own scope. Head completion rose from 205 to 298 of 436,
context_projection_unavailablefell to zero, and access control went from 0 to 52 completed questions. - Admission rejected genuine findings. The operation symbol was the text before the first
parenthesis, so
const preTax = eval(...)was the operationconst preTax = eval, which no qualified declaration could ever name. It is now the trailing qualified callee, with receiver spelling preserved. - Unresolved questions said why, and then there were none. The remaining 138 each named a function its file does not define. A route-registration module registers handlers other modules define, and regions were scoped to methods those files do not hold. The entry record keeps the handler name, because the authorization question is cross-file by nature, and the region is now scoped to the file instead. Every question the scan asks now completes on both revisions, 299 of 299, with detections unchanged.
Evidence: group 12 final and region attribution.
Verified
Immutable pushed snapshots, shared multi-ref deadlines, ranker-owned scope, physical vendor exclusion, compatible immutable graph/result reuse, actionability admission, compact truthful reporting and non-destructive hook integration are implemented. Advisory and blocking share the same evidence bar; optional model output cannot bypass it.
Contributor-scan task 2.19 is verified: the pinned frontend now retains the original Gruntfile, bringing the real NodeGoat census from 43/44 to 44/44 files. Partial, absent and malformed census no longer masquerade as a complete result or a falsely empty graph. The final code verification recorded 1321 passed tests, nine skipped, passing canonical gates and a packaged native PHP/JavaScript partition smoke. See owner verification.
Rollout blockers
- Useful checks within the deadline are unproven. The fresh post-census profile completed all 21 measurements: all 18 changed/cold/growth/multi-ref transactions timed out; the three identical-tip replays took 18.521–18.874s with valid graph/query reuse. Still 0/24 declared target checks completed. Maximum observed cancellation/report overrun was 0.874s, below the two-second allowance. After M1 establishes the useful comparison, amortize engine startup/loading and rerun the frozen seven-class profile; require warm p95 <=30s and >=95% completed supported checks.
- Target and change-context evidence remains incomplete. M1a measured exactly which contracts fail on the Node case: per-family change context, question-owned vendor relevance, operation correspondence for changed lines, file-wide correspondence ambiguity and query-to-operation provenance. VAmPI preserves raw findings at all three known sites, but the exact SQLI function question/transitive witness and supported authorization context are unresolved. Raw findings or rank positions cannot establish an actionable new regression. Repair this in the shared scan boundary and verify the actual associated query/path evidence.
- Independent quality qualification is missing. Freeze and review untouched Node API vulnerable/fixed/benign cases, including middleware/module/async context. Ghost remains reserved with missing reviewed inputs. PHP and Python need their own independent language/framework/family populations; Node results cannot qualify them. Require >=95% observed actionable precision with counts/intervals, >=90% supported recall, nonzero reviewed emitted alerts/positive controls and zero known fixed/benign false alerts.
- No current capability is eligible. The earlier registry is NO-GO and becomes stale under the repaired runtime/core identities. Loading it enables zero capabilities. After the preceding evidence passes together, regenerate exact-semantic eligibility and rerun packaged acceptance validation. Do not relabel old measurements as current.
C/libpng arithmetic and bounds remain explicitly unsupported; this is not a C/C++ detection qualification. Advisory rollout cannot be justified by silence from an unqualified registry.
Ownership and revalidation
Engine lifecycle, source/census and context/witness normalization belong to
contributor-scan. The hook consumes those contracts; no second engine, source IR,
ranker priority exceptions or differential graph mutation is introduced here. Graph
mutation would need a separate cold-equivalence experiment.
After owner fixes, refreeze identities and rerun ordinary scan gates, Python regressions, PHP-to-Node transfer, the representative cache/deadline profile, independent per-capability quality evaluation and packaged hook validation. Keep task 8.3 unchecked and the feature in implementation until those mandatory results are established.
Evidence: group 8 validation, post-census runtime scorecard, census repair.
The owner investigation now distinguishes missing family context producers, globally propagated vendor/correspondence gaps, and query-to-operation identity. See remediation research. These are next implementation prerequisites, not completed fixes or changed acceptance gates.
An isolated Joern session probe matched the full Node census with four warm requests in 0.221–0.467s and verified cancellation/reaping in 0.043s. It identifies a promising engine lifecycle change; full-query equivalence and changed-code performance remain unverified.
Family detection census, 2026-09-19
Four of seven declared families establish findings on the two development subjects:
injection, access_control, config_secrets and untrusted_destination. Output encoding, path,
prototype and deserialization have established nothing anywhere. Every question completed on
both subjects, 299 of 299 on NodeGoat and 35 of 35 on VAmPI, so the silences are answers
rather than failures — but three of them are diagnosed gaps rather than clean code:
NodeGoat's documented SSRF is invisible because needle.get is not in the sink vocabulary,
its XSS is invisible because output encoding declares only DOM sinks, and its two
access-control defects raise no obligation because the modelled data operations live in DAO
files while the family's scope is the handler's own file.
The census also found and fixed one false positive: NodeGoat's only prototype report was
the word set in a code comment, matched because the sink text clause searched a whole
assignment node. Nothing about this changes the release gates. No capability is qualified by
these subjects.
Two of the three gaps are closed. The untrusted-destination vocabulary now holds the client
surface a Node server actually calls, and NodeGoat's SSRF is reported at
app/routes/research.js:16 with a witness naming the call. Output encoding now asks about the
server half of its own family and establishes a finding at app/routes/research.js:25, and it
states in its declaration that template-mediated escaping is not something it can decide. Five
of seven families now establish findings.
Access control was the third gap and its structural half is closed. The obligation now follows the one named call a handler writes into its data-access module, attributed by the graph's own provenance rather than by name, and located at the call a contributor changed. On NodeGoat it is raised at six handlers where it was raised nowhere, and three reach a verdict. All three are corroborated and none is reported without a judge, which is the honest rung: entailing a carried obligation reported a signup handler that is public by design, because its authenticated siblings share the same data module. VAmPI is unchanged.
The commonest collection read now carries the obligation too, so NodeGoat's documented insecure
direct object reference raises one where it never had. find had been left out because the name
is shared with array searches and jQuery selectors, so it is claimed only in the argument shape
of a query document. Measured on a plugin that ships jQuery, all 88 of its find calls stay
unclaimed. Seven NodeGoat handlers now reach a verdict, all corroborated; the census is
unchanged, because the corroborated band needs a judge it does not configure.
Closing the second gap surfaced four defects, three of which were false positives it would
otherwise have shipped: a permissive-value set shared across settings whose booleans have
opposite polarity, a dotted setting matched on its trailing segment alone, a flow query asked of
a call's receiver as well as its arguments, and a file's settings ordered by the spelling of
their line numbers. Template auto-escaping disabled at server.js:135 is detected and still not
reported, because the arbiter returns one verdict per question and that file has an earlier
permissive setting; task 14.6 carries it.
Evidence: family detection census, untrusted destination vocabulary, output encoding family, access control across the module split, collection read obligation.