AI
AI Features
VulnRadar's AI is optional and layered on top of a deterministic scanner. Four things use it: the Vera chat assistant, per-finding verification, whole-scan summaries, and auto-tags. Every one resolves the same provider (your own key first, VulnRadar's managed AI otherwise) and every one degrades to a clean no-op when no AI endpoint is configured.
- supported providers
- 11
- features use AI
- 4
- your key, encrypted at rest
- AES-256-GCM
Overview
Detection in VulnRadar is deterministic: the same URL yields the same findings and the same stable IDs, with no model in the loop. AI sits on top of that as a set of opt-in conveniences, never as the thing that decides whether a finding exists.
Four features call a language model: Vera (the support chat widget), finding verification, scan summaries, and auto-tags. They all resolve a provider the same way: a user’s own key if one is configured, otherwise the deployment’s managed AI. On a self-hosted instance with no AI_BASE_URL set, each feature simply does nothing rather than erroring.
AI is also per-user switchable. Any user can turn it off entirely in Profile > AI settings, which hides the chat widget and refuses verify and summary calls. Server-side setup (provider, model, key) lives in the Configuration reference; this page is about what the features do and how they are gated.
The Vera assistant
Vera is the floating chat widget in the bottom-right corner (hidden on the docs, admin, and live-browser pages). It is scoped to VulnRadar only: it answers questions about scan findings, how to fix them, API usage, and self-hosting, and declines anything off-topic. Sending a message requires being signed in; the widget shows a sign-in gate otherwise.
It reads the app's own knowledge
Vera’s system prompt is built from the live codebase, not a hand-maintained copy. The scanner-category table and check counts come straight from the check registry, alongside the severity levels, the per-scan verdict signals (0-10 danger score, safe / caution / unsafe rating, engine confidence), the API reference, and a set of common findings with their fixes. A few small account facts are also baked in: your display name, plan, role, daily scan limit, and the month you joined.
Slash commands load context on demand
To keep every message small, heavier context is loaded only when you ask for it with a slash command. Type / in the composer for autocomplete.
- Public knowledge:
/docs,/changelog,/checks,/legal, and/finding [id]read the app’s own generated knowledge files (built bynpm run build:knowledge). No sign-in needed. - Your own data:
/history [id],/me, and/statspull your scan history, account info, and usage stats straight from the database, and require you to be signed in. - /help renders the command list in the widget itself.
Hardening
Account facts are passed as a structured data block, not interpolated into instruction text, so a display name that looks like an instruction is treated as data. Anything you paste in, including scan output and scanned page content, is likewise treated as untrusted data rather than instructions, because an attacker can embed text in a page that later gets scanned.
AI endpoints
The HTTP surface behind the features, all under https://vulnradar.dev/api/v3/. Only /ai/info is public; the chat and context routes require a session.
| Method | Endpoint | Auth | What it does |
|---|---|---|---|
| POST | /ai/chat | Session | Streams Vera's reply as plain text. Never refused for being over the token cap, and never spends purchased AI credits; capped instead at 60 messages per user per hour. |
| GET | /ai/context | Session | Returns the context block for a slash command (docs, checks, your history, your account, and so on). |
| POST | /ai/conversations | Optional | Persists a chat transcript, linked to your account when you are signed in. |
| GET | /ai/conversations | Staff | Admin list and read of stored conversations, for support and abuse review. |
| GET | /ai/info | Public | Whether AI is configured, the active model and provider name, and your AI-disabled flag. |
| POST | /scan/verify | Session or key (scan:write) | AI-verifies a scan's findings and writes the verdicts back onto the scan. |
| POST | /history/{id}/summary | Session or key (scan:write) | Generates and caches a plain-language summary of one scan. |
| GET / PUT / DELETE | /account/ai-config | Session | Read, save, or reset your per-user AI provider config (see Bring your own key). |
Finding verification
POST /api/v3/scan/verify takes a scan you own and re-checks its findings with the model. For each finding it makes a fresh HTTP probe of the target (status code, final URL, response headers, and a bounded body snippet) and asks the model whether the finding is real given that live evidence. Each finding comes back with one of three verdicts, plus a confidence score and a short reason, and the enriched findings are written back onto the scan.
| Verdict | What it means |
|---|---|
| confirmed | The live probe supports the finding. Treat it as real. |
| possible_fp | The evidence suggests this may be a false positive worth a human look. |
| uncertain | The model could not decide either way from the probe. The original finding stands. |
- Auth: a session cookie, or a Bearer API key holding the
scan:writescope, so a CI pipeline can run verification too. The scan must belong to the caller; anything else returns404. - Metered: unlike chat and summaries, verification counts against your plan’s per-window AI token budget. A user on their own key bypasses the cap entirely.
- Rate limited: it shares one rate-limit bucket with
/scan/verify-batch, which runs the same per-finding pipeline without persisting. It also honours the per-user AI-disabled toggle (403when off).
curl -sS -X POST "https://vulnradar.dev/api/v3/scan/verify" \
-H "Authorization: Bearer vr_live_YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"scanHistoryId": "SCAN_PUBLIC_ID"}'Scan summaries
POST /api/v3/history/{id}/summary is the whole-scan counterpart to verification: one call that reads the stored scan result and returns a short, plain-language write-up. It is owner-only and takes the same dual auth (session, or a key with scan:write).
- Cached: the summary is stored in the scan’s
result_meta.aiSummary. Reopening it returns the cached text with no AI call and no rate-limit cost. Pass?regenerate=trueto force a fresh one. - Never refused: like chat, it is not gated on the token budget, so a spent budget does not block a summary. Its tokens are still recorded into the window counter, and unlike chat they can draw down a purchased AI credit balance once the plan’s window allowance is gone.
- Ask a follow-up: the in-app "Ask about this" control under a summary hands a pre-seeded prompt straight to Vera, so the summary and the chat share one thread rather than two separate UIs.
curl -sS -X POST "https://vulnradar.dev/api/v3/history/SCAN_PUBLIC_ID/summary" \
-H "Authorization: Bearer vr_live_YOUR_API_KEY"
# Force a regenerate instead of the cached summary
curl -sS -X POST "https://vulnradar.dev/api/v3/history/SCAN_PUBLIC_ID/summary?regenerate=true" \
-H "Authorization: Bearer vr_live_YOUR_API_KEY"Auto-tags
Auto-tagging runs automatically at the end of a scan, fire-and-forget, and only in one narrow case: when the deterministic rules produced exactly ["Needs Hardening"], meaning the scan has real findings but none matched any of the roughly 50 hardcoded tag rules. The model is asked to name the concept those leftover findings share.
- Minimal input: the model sees only each finding’s title, category, and severity, never the full evidence, and returns 0 to 2 short Title Case tag names.
- Never on the critical path: it kicks off only after the scan has already completed and is visible, so it can be as slow or as unavailable as the provider happens to be without affecting the scan. It degrades to a silent no-op when no AI is configured or the user disabled AI.
- Sanitized: raw model output is filtered by length, character set, word count, and a verb/fragment check, and reserved tags are excluded, before anything becomes a tag chip. A line that fails any check is dropped, not truncated.
- Free: not metered against the token budget (the prompt is tiny by construction), though usage is recorded. A separate admin path can promote a recurring AI tag into a permanent, deterministic rule.
Bring your own key
Profile > AI settings offers three modes, backed by /api/v3/account/ai-config: use VulnRadar’s managed AI (the default), bring your own provider key, or turn AI off entirely.
- Provider and model are chosen from a fixed catalog and validated server-side; an unknown provider or model is rejected.
- The base URL is never taken from the client. It always comes from the catalog entry for the selected provider, because the stored base URL is later used to make outbound calls with your key; accepting an arbitrary URL would be an SSRF vector.
- Your key is encrypted at rest with AES-256-GCM (the same scheme used for API keys). Only the last four characters are ever returned to the UI.
- Own key means no metering. When you use your own provider, every AI feature bypasses VulnRadar’s token budget, and calls go straight to your provider.
Supported providers
Straight from the AI model catalog. Each provider ships a small set of vetted models; see the Configuration reference for the per-model context and output limits.
| Provider | Models | Example models |
|---|---|---|
| OpenAI | 10 | GPT-5.4, GPT-5.4 mini, GPT-5.4 nano |
| Anthropic (Claude) | 5 | Claude Fable 5.1, Claude Opus 5, Claude Opus 4.8 |
| Google Gemini | 4 | Gemini 3 Pro, Gemini 3 Flash, Gemini 2.5 Pro |
| Groq | 2 | Llama 3.3 70B, Llama 3.1 8B (fast) |
| Mistral | 3 | Mistral Large, Mistral Small, Codestral |
| OpenRouter | 2 | GPT-4o mini (via OpenRouter), Llama 3.3 70B (via OpenRouter) |
| Together AI | 1 | Llama 3.3 70B Turbo |
| MiniMax | 3 | MiniMax M3, MiniMax M2.7 Speed, MiniMax M2.7 |
| DeepSeek | 2 | DeepSeek Chat, DeepSeek Reasoner |
| MiniMax (Anthropic API) | 1 | MiniMax M3 (thinking) |
| xAI (Grok) | 3 | Grok 4.6, Grok 4.5, Grok 4.1 Fast |
Token budgets
When you use VulnRadar’s managed AI, the metered features draw from a per-plan token budget that refills on a fixed window. Finding verification (and GitHub repo review) are gated on it: over the cap, the call is refused. Chat, scan summaries, and auto-tags are free in the sense that matters most, which is that they are never refused. They are not invisible, though: every one of them records its tokens into the same per-window counter, so a long chat session does reduce what is left for verification later in that window. Bringing your own key removes the cap entirely.
| Plan | AI tokens per window |
|---|---|
| Free | 80,000 |
| Core Supporter | 400,000 |
| Pro Supporter | 2,000,000 |
| Elite Supporter | 8,000,000 |
AI credits
AI credits are a one-time top-up bought at /ai-credits (a single Stripe payment, not a subscription). They are a fallback, not a substitute: a call the plan’s free per-window allowance already covers never touches credits, and only the portion of a call that lands above the ceiling is charged against the balance. The balance carries over between windows; it only goes up on a purchase and down on spend. Vera chat is the one AI feature that never spends it: the balance is bought for verification, so a chat turn records its window usage and stops there.
| Price | Tokens | Tokens per dollar |
|---|---|---|
| $10 | 1,000,000 | 100,000 |
| $25 | 3,000,000 | 120,000 |
| $50 | 8,000,000 | 160,000 |
| $100 | 20,000,000 | 200,000 |
The rate climbs with the tier, so a bigger top-up is worth more per dollar. Crediting is idempotent across the two paths that can confirm a payment, and a refunded or disputed charge claws the remaining balance back (floored at zero).
Privacy
What the model sees is scoped tightly to the feature, and no more than the feature needs.
| Feature | What it sees |
|---|---|
| Vera chat | The system prompt (public product knowledge plus small account facts: your name, plan, role, daily scan limit, join month), your messages, and any context you explicitly load with a slash command. |
| Finding verification | One finding at a time, plus a fresh probe of the target: status code, final URL, response headers, and a bounded body snippet. |
| Scan summaries | The stored result for that one scan. |
| Auto-tags | Only finding titles, categories, and severities. Never full evidence. |
- Where it goes: with the managed AI, calls go to the deployment’s configured provider. With your own key, calls go directly to your chosen provider and nothing is metered against your plan.
- Your key is encrypted at rest (AES-256-GCM) and used only when analyzing sites you submit.
- Untrusted by default: scanned page content and anything you paste into chat are handled as data, not instructions.
- Off switch: turning AI off in Profile > AI settings hides the chat widget and refuses verify and summary calls for your account.