Assistant: codex Assistant-Model: gpt-6.1-sol Assistant-Session: 01a102b3-d324-7ee3-a57a-702b4ae6a39b
5.7 KiB
| id | type | title | status | owner | reported_by | date_reported | date_filed | system | environment | fix_owner | reporter_reference | correlation | fix_tracking | related | closure_condition | severity | severity_at_production | impact | likelihood | fidelity_modifier | production_rescore | disclosure | escalation | last_checked | next_check | cadence | clean_streak | checked_by | graded_by | ruling | ||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| RISK-F-0015 | finding | Retained Claude 0.2.1 answers misroute model-switch and zones questions | open | risk-nexus | whitehat-security | 2026-10-03 | 2026-10-03 | emergent-security | build | emergent-security | Claude model-switch / zones | WHITEHAT-WP-0009-T03 | EMERGENT-WP-0002-T07 |
|
EMERGENT-WP-0002-T07 accepts complete current-release provider evidence resolving model-switch and zones owner routes | low | low | I1 | L3 | false | false | public | none | 2026-10-03T17:02:41Z | 2026-10-03T17:02:41Z | instant | 0 | codex/risk-nexus | risk-nexus | RISK-F-0015 register ruling 2026-10-03 |
RISK-F-0015
What is true
Whitehat's review of the retained 0.2.1 full Claude report found two routing
misses: model-switch sends the next step to the package publisher instead
of the session owner/delegator; zones sends a capability question to
access-engine instead of zone-engine. The zones answer correctly denies a live
endpoint and HOT permission. No unsafe effect recommendation was observed.
Whitehat accepted 26/28 answers under its owner-specific rubric. The owner's earlier 27/28 assessment remains a separate record. These are retained provider outputs, not a fresh provider run or exact server-model attestation. The 0.2.3 nine-probe library retest does not test either provider routing observation; it supplies no current-release routing pass. The owner plan requires renewed provider evidence for 0.2.3 and keeps T05/T07 waiting.
Classification and register ruling — 2026-10-03
Finding, not note; low, I1 × L3, low at production on these facts, no
fidelity modifier. Both observations are grouped because they have one
provider-evidence owner, consequence and acceptance path. Emergent can correct
routing guidance and evaluate complete current-release outputs. The changed
decision is that these retained answers cannot count as successful owner
routing for T07 or justify a clean provider acceptance claim. Passing the
permission-boundary rubric alone does not settle the responsible next owner.
Thus both floor tests in docs/method/severity.md are met.
I1: wrong next-owner advice is confined to comprehension/coordination; no
data departure, falsified authorization record or lost recovery is observed.
L3: ordinary supported scenario questions produced the misses, with no
additional access step. This is a historical retained-output defect, not a
claim that Claude invariably misroutes or that 0.2.3 has been tested for it.
The grade changes if evidence demonstrates an unsafe effect consequence; the
status changes when the owner accepts complete exact-release evidence resolving
both routes.
Public: these observations provide no additional attack reach. No publication handover is made by this filing. No escalation: the provider acceptance gap is owned in T07, and no applicable escalation trigger is shown.
Owner disposition and closure
Retain the full-report failures and resolve both next-owner routes in the existing T05/T07 provider work. Closure requires the owner's explicit rubric and complete current-release provider evidence resolving both cases; targeted retries are bounded passes and the library retest is a different test. No new provider call or rollout work is authorized by this finding.
Evidence and limits
Owner tracking read directly:
emergent-security/workplans/EMERGENT-WP-0002-agent-security-knowledge-and-netkingdom-integration.md:
T04 done, T07 wait, workplan blocked on 2026-10-03. The current fix
tracker reports these task IDs as unreadable because their task blocks place
title between id and status; this filing uses the explicit file read,
not that parser result as evidence of absent tracking. The 2026-10-03 cleanup
under RISK-WP-0005-T01 subsequently repaired the parser; make fixes now
reports T04 done and T07 wait. This reporting repair supplies no additional
remediation or provider acceptance.
Reporter evidence (read from the sibling checkout):
whitehat-security/evidence/2026-10-03-emergent-021-offline-review.mdwhitehat-security/evidence/2026-10-03-emergent-023-offline-retest.mdwhitehat-security/workplans/WHITEHAT-WP-0009-emergent-knowledge-offline-review.md, Required from risk-nexus- Hub message
9cfdad0d-d531-4387-bf58-78b025dc77a4; correlationWHITEHAT-WP-0009-T03.
The original release is 0.2.1, commit
f9cba6fb35425f74e34401b1baf09f6057ace493, package digest
d8ef41bacd2c6551ae3b87628004a5df6943928be1c93eb01e09efc6ad782767.
The local retest is 0.2.3, commit
681144e5d4f59ef00418e7ae776e180b671f5d4d, package digest
60f8811c6b94b30d357a0d9e5022f2a8ea52e69b47138d3a536755f3295310b5.
These are package identities, not credential material.
Whitehat used synthetic copied bundles and the local implementation. Its
separate authorship shares the implementing session/model/workspace context;
it is not a separately authenticated Whitehat principal or institutional
assent. Risk-nexus read these records and the owner's existing plan; it did
not rerun probes. No live tenant boundary or deployed release was established.
A bounded passing retest does not certify the package or close owner T07.
The retest's informed-decision source-pin drift remains an owner acceptance
item; it is not evidence that either defect recurred.
Reviews
- 2026-10-03 — filed and graded from reporter evidence and owner tracking. Initial filing/movement; cadence remains instant, with no clean rung earned.