Skip to content
AI Security SentinelAI Security Sentinel

5. Adverserial AI — CyberKimi + CyberGLM

Contents7 numbered sections

Correction: v1 covered Lakera/Mindgard/Cisco/adversarial.com. The correct site is https://adverserial.ai/ (ADVERSERIAL with E), Adverserial AI LLC, Jersey City NJ. Two cybersecurity-fine-tuned models: CyberKimi (lordx64/cyberkimi) and CyberGLM (cyberglm), plus CyberSeek coming soon. Direct fetch 2026-09-28 of /, /docs.html, /founder.html below; benchmarks at github.com/lordx64/cyberkimi-benchmarks.

5.1 Identity

Section titled “5.1 Identity”
  • Legal: Adverserial AI LLC, a New Jersey LLC, 1078 Summit Ave #605, Jersey City NJ 07307, contact@adverserial.ai. Sources: https://adverserial.ai/terms, https://adverserial.ai/privacy (v Sep 16 2026).
  • Sites: https://adverserial.ai/, docs https://adverserial.ai/docs.html, founder https://adverserial.ai/founder.html, chat https://chat.adverserial.ai/ (OpenWebUI, auth-gated), billing https://billing.adverserial.ai/, terms /terms, privacy /privacy, waitlist /waitlist.html, API https://api.adverserial.ai/v1.
  • Founder: Taha Karim, Founder & Chief AI Researcher, creator of CyberKimi, handle @lordx64 — https://adverserial.ai/founder.html, links LinkedIn tahakarim, X @lordx64, HF lordx64. Bio claims: EPITA Paris cybersecurity; Symantec malware RE (1st CyberWar Challenge ~1,000); FireEye senior malware (co-author first LATENTBOT teardown); Head of Malware Research Labs DarkMatter (WindShift/G0112 WindTail/WindTape, HITB GSEC, ATT&CK S0466); Exodus Intelligence 0-day Android baseband chains; Google Play APK defense; now TikTok/ByteDance global threat intel AI workflows; Black Hat/HITB/SANS trainer/presenter. GitHub github.com/lordx64 (48 followers) corroborates LLC/US identity.
  • Team/date: Founding date and team size NOT published. Earliest artifacts Aug 2026 LinkedIn (“Meet CyberKimi”, CyberGym Aug 19). Public face solo founder + “team” mentions; Sep 2026 GitHub essentially lordx64 alone (168 commits/3 repos). Company LinkedIn ~37 followers (snippet). Do not cite team size.
  • NVIDIA Inception: Homepage footer shows “PROGRAM MEMBER / NVIDIA Inception” badge, but no NVIDIA listing found; only founder post “applied … forcing AWS quota now on NVIDIA Inception compute”. Treat as claimed-applied, unverified until NVIDIA directory confirms.
  • Positioning: “Intelligence. Engineered for cyber.” / “ENGINEERED FOR CYBER.” / “For authorized security work.” Two models fine-tuned for cybersecurity, red team + blue team, “Understand the adversary. Build a stronger defense.”

5.2 Models & Functionalities

Section titled “5.2 Models & Functionalities”
Model Base Params / Active Ctx Model ID Status
CyberKimi Kimi K3 + refusal ablation + cyber-tuning 2.78T / 104B active 1M total (750K client budget; 512K MXFP4 vLLM in bench) lordx64/cyberkimi Production, PAYG + chat
CyberGLM Undisclosed GLM family, cyber-focused — 131K client budget cyberglm In development, PAYG, actively refined
CyberSeek — — — — Coming soon, pricing TBA

Sources: homepage specs, https://adverserial.ai/docs.html FAQ (“builds on Kimi K3 with guardrails relaxed … + domain tuning”), GitHub (“Kimi K3 with refusal layer ablated and cyber-tuning, own GPU infra”), benchmarks (cyberkimi-v1 vLLM private node MXFP4 512K).

What it does: red team 01 + blue team 02 in one reasoning engine — detection engineering (Sigma/YARA/KQL), incident response (sequence reconstruction, containment/recovery), threat hunting (intel → testable hypotheses). Example workflow: “Review auth events … Identify sequence … propose detection strategy” → Evidence timeline + Detection logic + Response priorities. Domain specialization via ablation + cyber tuning from investigations/playbooks; privacy-by-design (no GPU request logs, no training on prompts); continuous dev toward RL in verifiable security envs.

Ablation nuance: founder’s separate phantom-kv repo describes “refusal removal as loadable KV-cache graft — zero weight modification, reversible”. Whether production CyberKimi uses phantom-kv vs weight ablation is not stated — do not conflate.

Caps: OpenAI-compat Chat Completions + Anthropic-dialect shim (sanitizes thinking blocks), streaming with reasoning_content/reasoning + content, prompt caching (reported in usage.prompt_tokens_details.cached_tokens / cache_read_input_tokens), tool calling via Kimi Code/Claude Code/Codex/OpenCode/Hermes configs (see §10.3). No weight download; Enterprise private-weight deployment on request.

5.3 Benchmarks & Evidence

Section titled “5.3 Benchmarks & Evidence”
  • CyberGym vuln reproduction: 86.7% (78/90, 65.6% first-attempt) — homepage. LinkedIn stratified 100-task subset 0.860 beating GLM-5.3 0.845, DeepSeek-V4-Pro 0.833, Gemini 3.5 Flash Cyber 0.832, Claude Mythos Preview 0.831, GPT-5.5 0.818, 2× stock Kimi K2.5 0.413, server-verified crashes, transcripts on request. Note 78/90 vs 100-task discrepancy = different runs/subsets — flag in proposals. Repo: https://github.com/lordx64/cyberkimi-benchmarks (14★, 3 forks, 0 watchers as of Sep 28).
  • ExploitBench V8 CVE-2024-6100: Methodology-assisted 10/16, Unassisted 8/16, Kimi K3 stock 4/16; single-seed; “not a universal ranking”. Method: bench-v8, 400-turn, seed 1, vLLM private node, stock CLI, per-episode transcripts — .../blob/main/CVE-2024-6100.md, leaderboard https://exploitbench.ai/env/v8-cve-2024-6100/ (Mythos 16.0 / GPT-5.5-Codex-AutoNudge 15.0 above assisted run).
  • Related founder repos: cyberkimi-pvp (129 commits, live CyberKimi vs models), phantom-kv (34 commits), pentestkit (“104/104 XBOW validation” claim), models.dev PR adding Adverserial as OpenAI-compat provider.
  • Skepticism to disclose: OffSeq Threat Radar Sep 5 (“No official CVE … independent verification lacking … launched with –no-sandbox”); Kobaran Sep 5 (“not independently verified … sandbox explicitly disabled … No firm/Chrome corroboration”); D. Kucinic Aug 13 (“0 refusals, ever … $149/mo … No ID verification … writes exploits as readily as detection rules” vs Daybreak gates); Phying aggregator “16/16 ACE” conflicts with repo’s own 8-10/16 — treat aggregator as unreliable.

5.4 Pricing

Section titled “5.4 Pricing”
Item CyberKimi CyberGLM
Input (cache miss) $8 / 1M $4 / 1M
Input (cache read) $0.80 / 1M $0.40 / 1M
Output (incl. reasoning) $30 / 1M $15 / 1M
Billing Prepaid wallet billing.adverserial.ai, top-ups $10/$25/$50/$100, auto-refill optional, real-time debit, $0 = stop Same wallet

Sources: homepage + https://adverserial.ai/docs.html pricing table. 1 credit = $1 at CyberKimi PAYG rates. Monero accepted per LinkedIn (Stripe alternative). Enterprise: dedicated capacity, private weights, SSO — contact@adverserial.ai.

Membership history (do not quote as current without qualifier): V1 Foothold $29/mo (12 credits/wk ≈$52 PAYG/mo, 10/day, 30K out/day, 100K/wk, 1 concurrent) / Hacker Manifesto $149/mo (60/wk ≈$260, 20/day, 150K/day, 600K/wk) / G0DM0D3 $349/mo (140/wk ≈$607, 40/day, 350K/day, 1.5M/wk, 2 parallel), ~42-44% off PAYG when fully used, weekly reset no rollover, memberships covered CyberKimi only. Docs now: “Memberships — Gone — legacy converted to wallet credit at full value”. Waitlist page: “stopping v1 … v1 beta concluded”, Priority Slot $149 one-time, Free $0, Enterprise trial — https://adverserial.ai/waitlist.html.

5.5 API

Section titled “5.5 API”
  • Base https://api.adverserial.ai/v1, Chat POST /v1/chat/completions, Models GET /v1/models, IDs lordx64/cyberkimi + cyberglm, Auth Bearer sk-.... Keys in billing dashboard, multi named keys (burp, ci-runner), independent revocation, wallet preflight; chat-platform API rejects end-user keys by design. Never embed in client-side/public repos. Errors: 401 revoked, 403 wrong endpoint, 402 empty wallet (/topup), finish_reason=length → raise budget, 503 retry w/ backoff.
  • Client budgets (conservative, not server cap): CyberKimi 750K (OpenCode example 65,536/8,192), CyberGLM 131,072. “Setting 1,048,576 does not increase server capacity.” Documented integrations: Claude Code/Cline (Anthropic shim https://api.adverserial.ai, 750K override), Kimi Code (~/.kimi/config.toml), Codex CLI ≥0.134 (Responses shim, provider+profile, 750K/700K compact), Codex desktop (key-file auth), OpenCode (@ai-sdk/openai-compatible, cyberkimi/lordx64/cyberkimi), Hermes (verified Sep 13; AWS_EC2_METADATA_DISABLED=true on non-AWS).
from openai import OpenAI
client = OpenAI(api_key="sk-YOUR-KEY", base_url="https://api.adverserial.ai/v1")
resp = client.chat.completions.create(model="lordx64/cyberkimi",
  messages=[{"role":"system","content":"You are a red-team operator assistant."},
            {"role":"user","content":"Write a Sigma rule for this behavior: ..."}],
  max_tokens=2048)

5.6 Legal / Privacy

Section titled “5.6 Legal / Privacy”
  • Terms (/terms): “Use only for lawful activity, with rights/authorizations required.” “Authorization comes first” — written auth (assets/techniques/windows/data-handling), stay in scope, bounty terms control. Permitted: lawful research, authorized red/blue, defensive engineering, RE, education, controlled testing. Prohibited: unauthorized access, malware/ransomware deployment, extortion/fraud/phishing outside engagement, disruption/sabotage, unlawful surveillance, IP theft, sanctions/export evasion, metering attacks. “Research label doesn’t legalize.” Human review mandatory before execution, isolated testing, allowlists/gates, no sole-AI life-safety decisions. “Do not assume guardrails capable of preventing unlawful activity. Response ≠ lawful. Capability never supplies permission.” Dual-use explicitly includes exploit analysis, vuln research, offensive techniques, malware analysis. Input rights retained (limited processing license, not to sell/train); Output assigned to customer where transferable; no weight ownership. AS-IS, $100-or-12mo-fees cap, indemnity, suspension, abuse contact@.
  • Privacy (/privacy + docs FAQ + homepage): “Privacy first: no inference/chat session logs, prompts not used to train.” “Your evidence stays yours. No GPU request logs.” Nuance: billing/usage retained (wallet, ledger, model, request ID, in/cached/out tokens — “not anonymous”, Stripe); “Temporary chat … not to save to history ≠ immediate erasure … runtime buffers and GPU prefix/KV caches remain until eviction”; metric samples 30-day cleanup; account/usage/payment do NOT share expiry; providers Heroku (app), RunPod (GPU), Stripe, Google sign-in; 30-day billing cookie; DPA/region/air-gap only by explicit agreement.
  • Net for consultancy: best contractual fit for authorized work (explicit auth-first + dual-use disclosure), but verify no-logging scope in writing for regulated clients (billing + KV-cache retention remains) and get DPA/region terms via Enterprise if needed.

5.7 Context: Defensive AI Scanners

Section titled “5.7 Context: Defensive AI Scanners”

Lakera (Check Point), Mindgard, Cisco AI Defense, Adversarial.com are defensive AI-app testing/guardrail platforms, not offensive LLMs — out of scope for this report’s three-generator comparison. Use only if a client needs AI-app audit pass-through ($100k+ enterprise); otherwise rely on the open-source regression tooling in §10.3 (Garak/PyRIT/Promptfoo).