21 Services from verdict.kechlab.com

Judges a (goal, action, untrusted-content) triple to detect prompt injection, goal drift, data exfiltration, and unauthorized payments before an AI agent acts, returning allow/review/block.

Given a screenshot and up to 8 UI element descriptions, returns click coordinates (fractional and pixel) plus visibility probability for each target element.

Evaluates whether a small model's answer is acceptable or needs escalation to a larger model by detecting refusal, evasion, off-topic responses, contradiction, and incompleteness.

Evaluates whether an AI agent should proceed autonomously, proceed with a check, or stop and ask a human before taking an action, based on irreversibility, ambiguity, missing info, and declared stakes.

Scans terms of service or contract text for risky clauses (auto-renewal, data resale, arbitration waiver, liability caps, unilateral changes, exit penalties, IP assignment) and returns a structured accept/review/reject recommendation.

Verifies whether a claim is supported, unsupported, contradicted, or uncertain relative to a reference source text, with probabilities.

Analyzes an MCP server or tool before installation to detect prompt injection, excessive permissions, lookalike names, and hidden instructions that could compromise an agent.

Scans untrusted text for prompt injection, data exfiltration attempts, secrets/API keys, PII, and toxicity before an LLM agent processes it

Performs semantic content moderation across categories including profanity, harassment, hate speech, sexual content, threats, self-harm, and scams, returning per-category scores and an allow/review/block decision.

Screens a customer-support bot's outbound reply for unauthorized promises, hallucinated facts, leaked data, off-topic responses, and tone violations before sending, returning send, human_review, or block.

Detects prompt injection and tool-poisoning attacks in tool/API results returned to an AI agent, returning allow, review, or block.

Analyzes a full trade history to determine statistical edge validity, optimal position sizing, edge decay detection, and funded account challenge compliance using rigorous quantitative methods

Detects supplier bank-detail change fraud (BEC) in payment emails by validating IBAN and sender domain against your supplier record and reading pressure signals, returning ok, verify_callback, or freeze_payment.

Validates a proposed trade order against user-defined risk rules and computes the maximum allowable position size before the order is sent to market.

Detects if a model reply has leaked secret canary tokens or paraphrased a confidential system prompt, returning clean, review, or leak.

Scans text for personally identifiable information (PII) using AI-level detection beyond regex, identifying names, addresses, IDs, health data, credentials, emails, IBANs, payment cards, and phone numbers — without rewriting the text.

Approves, reviews, or blocks shell commands proposed by coding agents using text rule checks and simulated execution, preventing dangerous operations like root deletion, key exfiltration, and reverse shells.

Scans an HTML page for hidden text invisible to humans (HTML comments, CSS-hidden elements, hidden attributes, Unicode-encoded content) and determines whether those hidden instructions are targeting an AI agent.

Analyzes text to return a sentiment score from -1 to +1 with confidence, a mixed-opinion flag, and per-aspect scores for named dimensions like price, support, or delivery — abstaining when text is ambiguous.

Classifies inbound leads into routing lanes (spam, job application, vendor pitch, existing customer, or real project) using hard rules and AI judgment to help sales agents triage and route inquiries.

Classifies an inbound customer message by category and urgency, scores it for spam, and returns a routing decision (queue, escalate, review, or discard).