Fine-Tuned Adapter and Merged-Artifact Security Scanning Before Deployment

You are the platform security engineer building pre-deployment security scanning for fine-tuning artifacts (LoRA/adapters, merged weights, tokenizer merges) in [INSERT ML PLATFORM NAME]. Artifacts originate from [INSERT SOURCES, e.g., internal teams, approved hub, vendors] and are destined for [INSERT SERVING CLUSTERS]; this is a defensive build task using synthetic and licensed test data only. Module 1 - Scope: define artifact types, size/format limits, required metadata (author, pinned base model, training-dataset hash, evaluation report, license), and the quarantine path. Module 2 - Static checks: unsafe-deserialization blocking, archive and permission inspection, embedded-code and custom-operator detection, secret scanning, and signature/provenance verification. Module 3 - Behavioral evaluation in sandbox: automated safety-benchmark deltas versus the base model, refusal-retention tests, canary-trigger probes, jailbreak-resistance regression suite [INSERT EVAL SET NAME/VERSION], and performance/cost bounds. Module 4 - Integrity checks: weight-diff statistics, adapter norm/scale outliers, and comparison against independently trained reference adapters to flag implausible modifications. Module 5 - Admission policy: specify policy-as-code rules (required sign-off, maximum tool permissions, environment tiers) and who may approve exceptions in [INSERT APPROVAL WORKFLOW TOOL]. Module 6 - Protection guidance: trust-tiered adapter registries, read-only mounts, runtime egress controls, and per-adapter credential isolation. Module 7 - Detection rules: telemetry for registry uploads from new identities, evaluation-score suppression attempts, and runtime behavior divergence; provide SIEM pseudocode and severities. Include a scanner architecture text diagram, CI pipeline integration steps, and abstract examples of failing versus safe artifacts (no exploit payloads). Provide the evaluation-harness configuration template and minimum coverage thresholds for risk tier [INSERT RISK TIER], plus a rollback and adapter-disable procedure with team notification. Map controls to [INSERT FRAMEWORKS, e.g., NIST AI RMF, ISO/IEC 42001, internal secure SDLC] and OWASP AI risk categories. List assumptions about existing build identity and registry infrastructure, and all prerequisites for the rollout. Define acceptance metrics (scan coverage, false-positive rate, mean time to block) and a quarterly rule-update cadence. Deliver the phased design plan first for approval, then scanner specifications, policies, and runbooks in sequence with review gates. ================