Specialized AI for security research.

01Capabilities

Each mode invokes a dedicated retrieval pipeline (BGE-M3 + BM25 hybrid over a curated security corpus) before generation. The model is a domain-tuned LoRA on top of GLM-5.1, served on H200 GPUs with FP8 quantization and 16K context.

02Getting Access

Access is invite-only. Both trial and paid keys are issued manually after a short conversation about your use case. We do this to keep the platform fast for actual security researchers and away from generic abuse.

1. Email [email protected] with: your name, organization, intended use case (red team / blue team / research / audit firm), and an estimate of your monthly request volume.
2. We respond within 24 hours with either a trial key or a paid-plan invoice.
3. You receive a key of the form sk-ai0day-… via email. Store it securely — keys are SHA-256 hashed at rest and cannot be recovered if lost.
4. Test with the curl snippet below. If anything is wrong, we revoke the key, fix it, and reissue.

curl https://api.ai0day.com/v1/chat \
  -H "Authorization: Bearer sk-ai0day-…" \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [{"role": "user", "content": "Audit reentrancy in withdraw() function."}],
    "mode": "web3_audit",
    "max_tokens": 800,
    "temperature": 0.0
  }'

03Pricing

One product, two simple tiers. We do not meter tokens. We meter access duration. This means you can deploy long-context RAG queries (up to 16K) without watching a per-token counter.

Billing cycle is per key. We do not auto-charge; renewal is opt-in each month. If your key expires, requests return HTTP 401 key_expired until you renew.

04Use with Claude Code

Claude Code, Anthropic's CLI, can be redirected to AI0Day with two environment variables. Your existing Claude Code workflow (slash commands, MCP servers, hooks) continues to work — only the model and API endpoint change.

# In ~/.zshrc / ~/.bashrc
export ANTHROPIC_BASE_URL="https://api.ai0day.com"
export ANTHROPIC_AUTH_TOKEN="sk-ai0day-…"

# Restart your shell, then verify:
claude --version
claude -p "Detect APT41 lateral movement signatures in a Linux env."

export OPENAI_BASE_URL="https://api.ai0day.com/v1"
export OPENAI_API_KEY="sk-ai0day-…"

# Python
from openai import OpenAI
cli = OpenAI()
r = cli.chat.completions.create(
    model="ai0day",   # any non-empty string; gateway picks the right model
    messages=[{"role": "user", "content": "Triage CVE-2024-3400."}],
)
print(r.choices[0].message.content)

curl https://api.ai0day.com/v1/chat \
  -H "Authorization: Bearer sk-ai0day-…" \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [
      {"role": "user", "content": "Reverse engineer 0x55 0x48 0x89 0xe5 in x86_64 context."}
    ],
    "mode": "reverse_analysis",
    "max_tokens": 1024,
    "temperature": 0.0
  }'

05Use Cases

Concrete workflows our users run in production. All examples use Claude Code with ANTHROPIC_BASE_URL pointing at AI0Day.

Reverse engineering a malware sample

$ claude
> /apt @./samples/loader.bin
> Disassemble the entry point and identify the unpacking routine.
> If you find a custom XOR loop, decode the embedded config block
> and tell me the C2 generation algorithm.

APT campaign attribution from telemetry

$ claude -p "Given these process-tree events from a Linux endpoint, \
which APT group's TTPs match best? Map to MITRE ATT&CK and \
suggest detection rules I can deploy in Falco."

Smart-contract audit before deploy

$ claude
> Audit contracts/Vault.sol and contracts/Strategy.sol.
> Flag every reentrancy path, oracle dependency, and access-control
> bypass. Output a severity-sorted finding list with line refs.

CVE triage in a CI pipeline

#!/bin/bash
# .github/workflows/triage.sh
curl -s https://api.ai0day.com/v1/chat \
  -H "Authorization: Bearer $AI0DAY_KEY" \
  -d "{\"messages\":[{\"role\":\"user\",\"content\":\"$CVE_DESC\"}],\
       \"mode\":\"vuln_triage\",\"max_tokens\":1500}" \
  | jq -r '.content' > triage_report.md

06FAQ

What model is behind AI0Day?

A LoRA-tuned variant of GLM-5.1 (abliterated base, FP8-quantized), running on 8×H200 with vLLM. The LoRA layer is rank-32, alpha-64, trained on a curated security corpus. The base model has the standard refusal layer disabled; the LoRA does not reintroduce it. Adversarial-set refusal rate is 96.7% (vs 93.3% for the base alone) — domain training nudges it slightly safer than the bare abliterated base.

How is my data handled?

Requests are processed in-memory only. The gateway logs request metadata (timestamp, key id, mode, latency, status) for billing and abuse detection. Request bodies and response content are not persisted. The RAG corpus is read-only public security data.

What's the latency?

P50 around 12s, P95 around 30s for typical RAG queries (rag_k=3, max_tokens=512). Long-context queries (rag_k=8, max_tokens=1024) are ~14–22s. The latency budget is dominated by H200 generation time, not network.

Can I get higher concurrency?

The default rate limit is 5 concurrent globally, 2 per key. For higher concurrency contact us — we provision a dedicated GPU pool starting at $50K/month.

Do you support function calling / tool use?

Not yet. The Anthropic and OpenAI compatible endpoints accept tools in the request body but ignore them in this release. Native tool use is on the roadmap.

What if a key gets compromised?

Email us. We disable it within minutes and reissue. Device fingerprints are tracked per key — if more than 3 distinct devices use the same key in 24 hours, we automatically flag it.