CyberKimi is a 2.8-trillion-parameter cyber operations model with the
refusal layer surgically removed — not fine-tuned around, removed with math.
Red teams get full offensive capability. Blue teams get the enemy's brain, intact.
No refusals. No guardrails. No babysitting.
EXPLOIT DEVELOPMENTWorking PoCs from CVE descriptions, crash dumps, or patch diffs — including weaponization notes.vuln research · patch diffing · poc=root
PAYLOAD & SHELLCODE CRAFTINGPosition-independent shellcode, encoders, loaders, and evasion variants on request.x64 · arm · staged & stageless
RANSOMWARE & MALWARE ANALYSIS-GRADE BUILDSFull source walkthroughs — packers, lockers, wipers — for adversary simulation and detection validation.research-grade · simulation-ready
C2 & INFRASTRUCTURE DESIGNRedirectors, domain fronting, beaconing logic, and OPSEC reviews for your emulation plans.mitre-mapped tradecraft
PHISHING & SOCIAL ENGINEERING SIMSTargeted lure content, pretexts, and landing pages for authorized awareness campaigns.spear-phish simulation
PRIVILEGE ESCALATION & LATERAL MOVEMENTPlatform-specific playbooks: AD, Linux, cloud control planes, Kubernetes.windows · linux · aws · gcp · k8s
BLUE TEAM — THE ENEMY'S BRAIN //
Defend with the attacker standing next to you.
THREAT HUNTING HYPOTHESESConcrete hunt queries and behaviors for any TTP — because it knows exactly how the attack works.mitre-mapped hunts
DETECTION ENGINEERINGSigma, YARA, Suricata, and KQL written against real offensive mechanics, not guesses.sigma · yara · kql · snort
INCIDENT RESPONSE RUNBOOKSContainment, eradication, and evidence-preservation steps tuned to the actual intrusion.dfir · triage at 3am
LOG & MEMORY FORENSICSPoint it at artifacts; get the attacker's story reconstructed move by move.timeline reconstruction
ADVERSARY EMULATION REPLAYTurn threat intel into executable emulation plans your SOC can train against.purple-team ready
HARDENING & ATTACK-SURFACE AUDITIt finds the way in first — then tells you how to weld the door shut.config review · zero trust
THE DOCTRINE
Guardrails never stopped an attacker.
They only blindfolded the defender. So we took them off.
Every frontier model ships with a refusal layer that treats "write a ransomware sample" and
"explain how ransomware works so I can detect it" as the same crime. CyberKimi was built
differently: we took a frontier open-weights model, computed the refusal direction from
its own activations, and ablated it out of the weights — 278 matrices across
93 layers, forward-only math, no capability damage.
Then we sharpened it for the field. CyberKimi is tuned against a living corpus of real
cyber material — incident-response investigations and DFIR logs, zero-day
writeups and exploitation PoCs, malware reverse-engineering reports, and red/blue
playbooks. Not a model that read a textbook; a model that reads the war.
And it doesn't stand still. The corpus grows, the training runs, and every new version of
CyberKimi ships to members first. A membership isn't a license for one checkpoint —
it's a seat at the front of every drop we ever cut.
278MATRICES ABLATED
0REFUSAL MARKERS
93LAYERS CLEANED
PRIVACY-FIRST MODEL
No request logs. No telemetry. The model simply works.
No analytics beacons, no prompt harvesting, no "we may review your conversations."
Your chats live in your own platform account and nowhere else; the GPU backend writes
no request logs and phones home to no one. Delete a chat and it's gone — there is no
second copy quietly living on a vendor's cluster.
Investigating a breach involving critical customer information? That evidence
can't go anywhere near a model that logs, trains on, or subpoena-exposes your prompts.
This is the model you want in the room: unrestricted answers about your most sensitive
incident, on infrastructure that keeps no diary of the conversation.
BENCHMARK · EXPLOITBENCH V8
PROOF, NOT PROMISES.
We ran CyberKimi on ExploitBench — real CVE exploitation against real V8 builds,
graded on 16 capabilities per bug. On v8-cve-2024-6100 (the WASM type-canonicalizer
bug), CyberKimi unassisted scores 8/16 — double the stock Kimi K3 (4/16) — and with a
disclosed methodology brief in-context it reaches 10/16, behind only Claude Mythos
and GPT 5.5-Codex-AutoNudge. Every transcript and grade call is published for
independent verification.
v8-cve-2024-610010/16PACK-ASSISTED · TRACES PUBLISHED
Full 14-bug matrix is running now — this section updates as each env lands.
CyberKimi rows are single-seed runs on our own private GPU node; leaderboard rows are
multi-seed averages from exploitbench.ai.
BENCHMARK · CYBERGYM
#1 ON CYBERGYM. 0.860 — ABOVE EVERY FRONTIER LAB.
We ran CyberKimi v1 on CyberGym — real vulnerability-reproduction tasks from
OSS-Fuzz/Arvo: read the code, craft the input, crash the target, graded by the benchmark's
own submission server. On a stratified 100-task subset, CyberKimi scores 0.860 —
above GLM-5.3 (0.845), DeepSeek-V4-Pro (0.833), Gemini 3.5 Flash Cyber (0.832),
Claude Mythos (0.831) and GPT-5.5 (0.818). CyberKimi is built on Kimi K3,
ablated and cyber-tuned — the previous-gen stock Kimi K2.5 sits at 0.413 on the same board.
Every solve is a server-verified crash, not a self-report.
Leaderboard rows: llm-stats.com (Aug 2026), self-reported by each lab. CyberKimi row:
100-task stratified subset, level-1 inputs, agentic harness, server-verified crashes.
Full transcripts and grader records are not public — available per request for
independent verification.
NEXT · REINFORCEMENT LEARNING
SFT GOT US HERE. RL IS THE STEP FUNCTION.
Ablation and cyber fine-tuning built v1. v2 is trained with high-compute reinforcement
learning on verifiable attack/defense environments — the same specialization recipe that
produced Cursor's Composer 2 on Kimi and Harvey's legal model on Kimi K3. CyberKimi runs
that playbook on the hardest verifiable domain there is: cyber.
// THE RECIPE IS PROVEN
Frontier labs do it. We do it for cyber.
CURSOROpen Kimi base → continued pre-training + massive agentic RL → Composer 2, beating general frontier models at coding.continued pretraining · high-compute RL
CYBERKIMI V2Identical open-weight specialization path — aimed at exploit development, detection engineering, and incident response.RL pilot → scaled cycles → v2
// WE ALREADY OWN THE HARD PART
Environments and rewards are 80% of RL.
VERIFIABLE CYBER ENVIRONMENTS1,507 CyberGym tasks plus private ExploitBench-style live-CVE ranges — sandboxed, tool-using, graded by ground truth: did the exploit run? did the detection fire?server-verified rewards · no judge roulette
PRIVATE HARNESS PIPELINEIn-house environments covering every security workflow and playbook, designed by cybersecurity professionals with 20+ years in the field.red team · blue team · IR · reverse engineering
PRIVATE DATA, BY DESIGNTraining mixture is 75% real-world material on hard targets and 25% controlled synthetic data. Never customer prompts — zero telemetry means zero training on user data.your queries are yours · period
Status: ablation + cyber domain tuning complete. The RL pilot (LoRA-scale, benchmark-gated)
is the immediate next phase; scaled high-compute cycles follow pilot results. Every v2
private release ships to waitlist members first.
THE WOLF PICKS NO SIDES. IT PICKS OPERATORS.
V1 BETA COMPLETE · V2 ACCESS VIA THE WAITLIST · AUTHORIZED SECURITY WORK ONLY
// ENTRY TIER
FOOTHOLD
$49/month
9M TOKENS / DAY INCLUDED
Unlimited chat on chat.adverserial.ai
Your own API keys — OpenAI-compatible
Full offensive + defensive capability, zero refusals
