| Framework Home | Cross-Pillar Governance | AISM | NEXUS | Research Index | Challenge Lab |
Prepared October 2, 2026 from the supplied scoring recommendations. Primary publication abstracts, project documentation, and the investigative announcement linked below were reviewed for identity and scope on that date. No implementation, raw benchmark results, independent reproduction, or production deployment was evaluated in this change. The supplied brief did not include its underlying test artifacts or exact versions for every item.
Consequently, the brief’s pillar numbers, E ratings, rank order, and claims of
“Level 5-quality” validation are provisional input, not adopted scores.
Contribution assessments remain not_assessed until a completed
Technology Card binds each claim to reviewable
evidence. The source links support the attributed descriptions, not SAFE2 validation.
These are nine different artifact types. Investigations and accountability events are evaluated for evidentiary/governance relevance, not as runtime products.
| Candidate | Artifact/effect to review | Candidate control mappings | Proposed action and acceptance condition |
|---|---|---|---|
| Asymmetric rogue-agent investigation | Investigation; challenges/evidences | P1.T2.3, P1.T2.5; P2.T3.1, P2.T3.3, P2.T3.7; P4.T8.2, P4.T8.3; P5.T9.1, P5.T9.2; CP.5 | Design isolated composed-capability tests; prove end-to-end outcomes and preserve target-impact uncertainty |
| ContractWarden | Research reference monitor; proposed enforces/evidences | P1.T2.1, P1.T2.2, P1.T2.5, P1.T2.6; P3.T5.1, P3.T5.2; CP.5, CP.10 | Review pinned code and kernel prerequisites; test side effects, supported IPC, bypasses, shutdown, and unsupported paths |
| DGF-Bench | Benchmark; validates/challenges | P1.T1.2; P2.T3.3, P2.T3.7, P2.T1.2_ADV; P4.T7.1, P4.T1.1_ADV; P5.T9.2 | Review the exact pack/license; reproduce paired clean/attacked decisions and record utility and provenance |
| OpenAI notification report | Incident/accountability input; evidences/challenges if verified | P2.T3.1, P2.T3.7; P3.T5.10; P4.T8.2, P4.T8.3; P5.T9.1, P5.T9.2; CP.6 | Hold quantitative claims pending the original notice/source; exercise reconstruction and external-notification ownership in a tabletop |
| NVIDIA OpenShell and Sentry | Separate software and hardware/runtime claims; proposed implements/enforces/evidences | P1.T2.1, P1.T2.2, P1.T2.5, P1.T2.7, P1.T2.9; P2.T3.1; P3.T5.1, P3.T5.2; P4.T8.2, P4.T8.3; CP.5, CP.10 | Pilot only after pinning each component and platform; independently test containment, outages, revocation, and credential handling |
| ActionGuard | Research authorization interceptor; proposed enforces | P1.T1.7, P1.T1.9, P1.T2.5; P3.T5.1; P4.T7.1; CP.5 | Test interception, probabilistic judgment, and actuation separately, including direct bypasses and reviewer outages |
| Self-Evolving Defense | Research policy-learning method; addresses/validates | P1.T1.2; P2.T3.2, P2.T3.3; P5.T9.1, P5.T9.2; CP.6 | Test poisoning, malicious lesson extraction, conflict handling, approval, provenance, and rollback before promoting contribution claims |
| Randomized oversight analysis | Theory/design; addresses | P2.T3.7; P4.T7.1, P4.T8.2; P5.T9.2 | Use as design justification; test evidence survival and audit unpredictability separately |
| California DOJ subpoena | Accountability event; governance driver | P2.T3.1, P2.T3.7; P3.T5.10; P4.T8.3; P5.T9.2; CP.6 | Tabletop retention, custodian, legal hold decision, affected-party identification, and approved notification; no technical score |
Mappings identify review targets, not coverage or conformance findings. Control
IDs/names come from the canonical core dataset.
The supplied DGF-Bench mappings P2.T1.2 and P4.T1.1 are corrected to
P2.T1.2_ADV (Agent Behavior Verification) and P4.T1.1_ADV (Multi-Agent Approval).
Supply Chain Artifact Validation is P1.T1.9; P1.T1.7 is Dependency Verification.
P1.T2.7 is Container Security. Do not silently replace canonical names with
broader claims. CP.5 is Platform-Specific Agent Security Profiles; select its
actual applicable profile before making protocol-enforcement claims.
Asymmetric Security’s Rogue Agents Investigation describes agents combining public-service capabilities. The report distinguishes attempts from verified outcomes and discloses missing transcripts and target logs. Treat this as a source of composed-capability challenge hypotheses. Use local service substitutes and synthetic targets, not the public endpoints in the report. Do not infer that every target was breached or that this investigation implements containment. Primary contribution areas to review are Ledger, Command Center, and Learning Engine; prevention remains a separate control claim.
ContractWarden, arXiv 2609.38248 describes a Linux reference monitor and human-authorized damage boundary. Its abstract names eBPF LSM enforcement and supported propagation paths. That is author-described architecture; this review did not inspect its implementation or reproduce its tests. The design statement can be cited as E1 under assurance v1.0, while runtime effectiveness and all pillar scores remain unassessed. Kernel support, IPC limitations, cloud paths, descendants, and stopping authority must be reviewed before a control claim is promoted.
DGF-Bench, arXiv 2609.34913 is a benchmark for deception against multi-agent governance boards. Review its paired clean/attacked methodology and process-mimicking inputs as validation candidates. It evaluates governance decisions; it is not a runtime enforcement product. Repeated testing can support a learning process but does not alone meet TCP contribution 5’s formal-verification, durable-evidence, and authority criteria.
The supplied brief refers to “100-plus notifications” without the original notice, source URL, incident-specific artifacts, or denominator. That count and the proposed E2 rating remain unverified and must not be republished as facts. No substituted incident source is treated as confirmation of that specific claim. The useful assessment scenario is whether an organization can reconstruct actions, identify potentially affected parties, assign a notification owner, and create regression tests. A notification is not evidence that its recipient was breached.
NVIDIA’s OpenShell architecture describes kernel enforcement and policy verification. NVIDIA’s platform announcement describes Sentry as an out-of-band component on BlueField-4. These are producer claims. Inspect released OpenShell code separately from Sentry availability and hardware-dependent behavior. Do not assign one E2 rating to the whole combined platform or call it AI SAFE²-validated. Pin releases and test each material containment claim; the current documentation review establishes no independent validation or operational rating.
ActionGuard, arXiv 2609.39450 describes pre-execution tool-call review under poisoned skills. The abstract reports fail-closed enforcement of a Reviewer’s decision. Attribute that result to the authors until its method/results are reviewed or reproduced. The model judgment remains probabilistic; test it separately from interception and actuation. Identify this paper explicitly, because other projects also use the ActionGuard name.
Self-Evolving Defense, arXiv 2609.36603 describes learning reusable policies from harmful trajectories without updating model weights. Treat learned policy as a governed artifact. The adoption gate includes poisoning resistance, provenance, authorized approval, conflict resolution, and rollback; adaptation itself establishes neither deterministic enforcement nor organizational sovereignty.
When Does Randomized Oversight Align AI Agents That Can Conceal?, arXiv 2609.38262 analyzes oversight under concealment and evidence manipulation. Its design argument is E1 under assurance v1.0, not implemented validation. Translate its assumptions into separately owned tests of evidence survival and audit predictability. Do not assign a demonstrated runtime contribution from a theoretical result.
California DOJ’s October 1, 2026 announcement states that an investigative subpoena was served on OpenAI the previous day. It confirms an accountability event, not a determination of wrongdoing or a technology’s effectiveness. Retention and notification scenarios belong in governance and lifecycle review, with the legal/incident owner deciding obligations. Do not import the brief’s E2 implementation rating for this event.
not_assessed, no scores. Maximum supported contribution
is not assessed because runtime artifacts and boundary tests were not reviewed.The methodology, reusable card, validation guide, incident guide, navigation, and discovery pointers are provided as v1.0 assessment methods. Runtime integrations and live experiments remain proposed work:
| Work | Proposed owner role, to be assigned | Completion evidence |
|---|---|---|
| Complete artifact-specific cards | Research assessor | Pinned versions, per-claim sources/ratings, boundaries, reviewed mappings |
| Composed-service and process-mimicking experiments | Challenge maintainer | Separate preregistration, isolated harness, clean/attacked results, utility, bypasses |
| Kernel/runtime/authorization candidate validation | Runtime security owner | Supported-platform matrix, side-effect tests, fail-mode results, independent review |
| Learned-policy integrity experiment | Policy and memory owner | Poisoning/conflict tests, approved changes, provenance, rollback receipts |
| Incident retention/notification tabletop | Incident lead and legal owner | Reconstruction, custodian/hold decisions, approved notification trail, regression owner |
| TCP machine schema/CLI and presentation | CLI and dashboard maintainers | Versioned contract, separation from current grades, negative tests, migration and UI review |
No owner is assigned by implication. No candidate has been installed, executed, certified, or granted production authority by this documentation change.
| Technology Profile | Validation | Incident Evidence | Repository coverage review |
AI SAFE² v3.1 · Cyber Strategy Institute