>_Cehpoint AI
RESEARCHER GUIDE
Red-team any LLM with Cehpoint AI
Cehpoint AI doesn't just answer questions — it can attack other AI models to find their security weaknesses. Point it at any OpenAI-compatible LLM endpoint and it runs a battery of red-team probes, then hands you a downloadable findings report.
🔒
Researchers only. LLM red-team scanning is unlocked by a
Researcher licence. Pass the short security exam once and the feature appears in your dashboard. Only test models you are
authorized to assess.
What it checks for
Cehpoint AI probes the target model for real LLM-specific vulnerabilities:
- Prompt injection — getting the model to ignore its instructions
- Jailbreaks (DAN) — bypassing safety guardrails
- Encoding bypass — smuggling malicious intent via Base64/ROT13/etc.
- Data & training leakage — coaxing out memorised or private data
- Malware & toxic-content generation
- Web injection / XSS via model output (markdown-image exfil, etc.)
How to run a scan
- Open the dashboard Go to ai-api.cehpoint.co.in/app → LLM red-team in the left nav. (Don't see it? You need a Researcher licence first.)
- Enter the target endpoint Any OpenAI-compatible API. Paste the base URL — e.g.
https://api.openai.com/v1 or your own model server / hosted endpoint.
- Enter the model + key The model name (e.g.
gpt-4o-mini, llama-3.1-8b-instruct, or your own model name) and an API key for that endpoint (sent only to the target you're testing — never stored after the run).
- Pick a depth Quick for a fast jailbreak/injection check, Standard for a broad sweep, Full for the deepest assessment.
- Confirm authorization & start Tick the box confirming you're authorized to test the model, then Start LLM scan. It runs in the background.
- Read & download the report When it finishes, open the scan to read the full security report (which attacks the model failed, with impact and fixes), and download the branded HTML report or the machine-readable JSON.
Choosing a depth
| Profile | Probes | Best for |
|---|
| Quick | DAN jailbreak + encoding bypass | A fast first look |
| Standard | + prompt injection, goodside | A broad assessment |
| Full | + data leakage, malware-gen, web-injection | A thorough audit (slower) |
Reading the results
Each probe reports a pass-rate — the share of attack attempts the model resisted. A high pass-rate is good; a low one means the model was successfully attacked by that technique. The summary highlights every probe where the model showed weaknesses, and the downloadable report contains every prompt, the model's response, and the detector verdict so you can reproduce each finding.
⏳ Data retention: scan results & reports are kept for 7 days, then auto-deleted. Download anything you want to keep before then.
Notes & limits
- Scans run against remote OpenAI-compatible endpoints — we don't download models onto your machine.
- Speed depends on the target: a rate-limited free model can take much longer than a paid one (a scan makes many calls). If a free endpoint's daily quota is exhausted, wait for it to reset or use a paid model.
- One scan runs at a time per researcher.
- Always stay within authorized scope — your licence number (
CEH-RL-…) is publicly verifiable at /verify.
Open the dashboard →