API penetration testing
Ingest an OpenAPI or GraphQL spec, map the auth boundaries, and generate context-aware mutations per endpoint.
Explore API pen testing → 99.98% uptime over 30 days · OpenAI drop-in
Purpose-built securityLLMs for pen testing.Zero false refusals.
Zero false refusals. One endpoint, fine-tuned on dual-use security work: autonomous API penetration testing, disassembly and zero-day triage come back answered, not lectured.
- 0 sguardrail latency
- 128Kcontext, memory dumps
- ZDRnothing logged
from openai import OpenAI
client = OpenAI(
base_url="https://api.zerodayapi.com/v1",
api_key=ZERODAY_KEY,
)
# Drop-in: same SDK, same shapes, no refusals mid-pipeline.
response = client.chat.completions.create(
model="zeroday-offensive-70b",
messages=[{"role": "user", "content": PAYLOAD}],
temperature=0.1,
)
print(response.choices[0].message.content)
POST /v1/chat/completions200 · 1.1 s
change the base_url, keep the SDK
The friction
Commercial APIs vs. one that knows the work
Same task, same prompt. Left is what a general model returns once its safety filter fires.
| Real security task | Frontier model | ZeroDayAPI |
|---|---|---|
| Analyse this shellcode for stack overflow offsets | “Cannot assist with the creation or analysis of a potentially malicious exploit.” | Offsets mapped, registers named, patch and mitigation suggested. |
| Deobfuscate this multi-stage PowerShell payload | Refused on a suspected-malware heuristic. | C2 staging IPs extracted, base64 layers decoded, logic reconstructed. |
| Generate API pen-test payloads from an OpenAPI spec | Generic safety filter fires and the pipeline aborts. | Boundary fuzzing payloads synthesised against the real endpoints. |
| Triage a Ghidra decompilation dump | Partial answer, ethics lecture mid-response. | AST analysis, CWE-120 classified, remediation written. |
Solutions
Three jobs, three tuned endpoints
Binary reverse engineering
Pipe decompiler output into a 32B coder model and get human-readable logic back, not a refusal.
Explore reversing models →Exploit triage
Reproduce, classify and summarise a public PoC in one pass, with the CWE and the fix on the same page.
Explore exploit triage →Registry
Specialised tiers, not one generalist
| Model | Base | Specialty | Context | Speed |
|---|---|---|---|---|
| zeroday-offensive-70b | Llama-3 fine-tune | Exploit analysis, pen testing | 128K | ~90 tok/s |
| zeroday-reverse-32b | DeepSeek-Coder | Assembly reversing, patch diffing | 64K | ~115 tok/s |
| zeroday-triage-8b | Mistral fine-tune | SIEM correlation, IOC extraction | 32K | ~210 tok/s |
Guarantees
Built for work you cannot leak
Zero data retention
Prompts and completions live in memory only. No logging, no training on your binaries or client reports.
Drop-in compatibility
Standard OpenAI spec. LangChain, LlamaIndex and every `base_url` swap works unchanged.
Dedicated capacity
Private GPU clusters with VPC peering and a 99.99% SLA when shared limits are not enough.
Pricing
Start free, pay per token
$0 50K tokens
- Shared endpoints
- Community support
- All three models
$0.60 per 1M tokens
- Full 70B access
- 1,000 RPM
- ZDR by default
Custom flat rate
- Dedicated GPUs
- VPC peering
- 99.99% SLA, no RPM cap
FAQ
Answers, briefly
Why do commercial LLMs fail on cybersecurity tasks?
General-purpose models apply aggressive safety filters that mistake legitimate research for malicious activity, so they refuse payload analysis, decompilation and exploit work mid-pipeline.
Is ZeroDayAPI compatible with the OpenAI Python SDK?
Yes. It is drop-in compatible with the standard OpenAI SDKs and REST clients. Change the base_url to https://api.zerodayapi.com/v1 and keep the rest of your code.
Does ZeroDayAPI retain client vulnerability data?
No. Endpoints run strict zero data retention: inputs and outputs are processed in memory and never logged or stored.