7. Legal Workflow

7.0 Ground rules

All three require explicit written authorization per target, scope-of-work, dates, IP ranges, accounts, forbidden actions (prod data exfiltration, persistence, lateral movement beyond scope), data handling, retest window. No authorization = ToS violation + criminal exposure. Mirror vendor disclaimers in your MSA.

7.1 OrcaRouter in workflow

1. Pre-engagement: apply for orca/orcacyber-zero-1.0 with engagement letter (component, vuln class, quarter). Do not script around refusal/failover to evade vetting.
2. Scoping: create workspace per client, keys scoped to orca/orcacyber-zero-1.0 + named dual-use bases (openai/gpt-5.5, anthropic/claude-opus-4.8 for report writing — Zero’s vendors note cyber models write reports worse), spending caps, orcarouter/auto disabled for scope control.
3. Execution: vuln triage/repro across large repos (1M ctx), root-cause, red-team tooling gen, patch validation. Log every call (Requests + x-orca-resolved-model receipts). Enforce ‘evaluation before reliance’ (Terms §6).
4. Deliverable: attach token receipts, resolved-model IDs, cache HIT/MISS, guardrail triggers for audit defensibility. Use orcaverify-text1.0 to check AI-assisted phishing samples if relevant.
5. Post: revoke keys, retain logs per retention policy, invoice pass-through $3/$5 + Team overhead.

7.2 Abliteration in workflow

1. Pre-engagement: Developer $20 key for lab, Growth $50 for team audit logs, Scale $200 ($200 credit included) for heavy ExploitGym runs. Do NOT use anon free-tier for client data (1 call, images only).
2. Scoping: per-client projects, scoped ak_ keys, per-user/project quotas, independent revocation, SIEM via Policy Gateway (Enterprise — budget this if client requires EU Act evidence; otherwise export token counts/timestamps/model IDs manually).
3. Execution: CVE repro, exploit dev, malware analysis (not delivery), jailbreak/prompt-injection suites (direct PDF/email RAG exfiltration, tool-misuse coercion), synthetic phishing/harassment datasets for detector tuning (paid 3-row preview → HF/S3 export). Use abliterated-model-large-v2 max reasoning for Exploits, abliterated-model base for multimodal (screenshots). Disable web search/fetch for client-confidential prompts (third-party retention voids zero-retention).
4. Deliverable: include authorization ID, model ID (large-v2 GLM-5.3 FP8), reasoning mode, flagged_categories, policy decision/reason if Gateway used, plus human validation statement (vendor outputs may be inaccurate).
5. Post: delete generated training datasets if client requires, retain billing metadata only, suspend keys. Note ToS gap to client: vendor allows authorized testing but bans illegal/harmful/high-risk — your letter is the shield.

7.3 Adverserial in workflow

1. Pre-engagement: top up wallet ($25-50 reserve, Monero option if needed), create named keys per tool in billing dashboard; note V2 waitlist for chat, API is wallet-direct.
2. Scoping: per-client keys, independent revocation, wallet preflight ($0 = stop). Map to OWASP / ATLAS for report structure. Get written auth per Terms S04.
3. Execution: hard vuln repro, V8/CVE analysis (cite public transcripts as precedent), Sigma/YARA/KQL, IR timeline, threat-hunt hypotheses. CyberKimi for depth, CyberGLM for cheaper drafts. Streaming + reasoning fields for audit trail; isolated testing + human review before execution.
4. Deliverable: include model ID, cache reads, benchmark context (86.7% subset, 10/16 assisted V8 single-seed + skepticism notes), privacy scope (billing/KV-cache retained), plus human validation (AS-IS, $100/12mo cap).
5. Post: revoke keys, retain billing ledger, invoice PAYG + Enterprise uplift if needed. Add Garak/PyRIT/Promptfoo regression between engagements.