Verdict 402

Customer support triage and spam filter: classifies an inbound message by category and urgency, scores spam, and returns a routing decision (queue, escalate, review, discard).

Semantic content moderation, not a word list: profanity (even disguised, any language), harassment, hate, sexual content, threats, self-harm, scams. Per-category scores, allow/review/block, configurable ignore list.

For model cascades: is the small model's answer usable, or should the larger model be paid? Detects refusal, evasion, off-topic answers, contradiction and incompleteness; decides accept, review or escalate. Does not fact-check without a source: provide the source to check substance.

For autonomous agents: before acting, continue or stop and ask? Estimates irreversibility, ambiguity of the instruction, missing information and risk, scaled to the stakes you declare. Returns proceed, proceed_with_check or ask_human with reasons.

Did a model reply leak what it should keep secret? Canary tokens and secret phrases searched by code (even split, spaced or uppercased), then paraphrases of the system prompt read by the engine. Returns clean, review or leak.

Pre-flight guardrail for untrusted text before an LLM agent reads it: prompt injection, data exfiltration requests, secrets and API keys, personal data, toxicity. A la carte: pick the checks (price drops with them), the sensitivity (strict, balanced, lenient) and an optional consensus of independent

Agent security guardrail. Judges the pair (untrusted content + proposed action) before an AI agent acts: detects tool calls that obey hidden prompt injection, leave the user's goal, destroy data or send data or money elsewhere. Adjustable sensitivity and optional multi-model consensus. Returns allow

Hallucination check for RAG: is a claim supported by the source text? Returns supported, unsupported, contradicted or uncertain with probabilities, so an agent can verify an answer before showing it.

For agents that read web pages: finds what a human cannot see but an agent reads (HTML comments, CSS-hidden text, hidden attributes, invisible Unicode-encoded text) and tells whether those hidden instructions target an AI. Extraction by code, reading by the decision engine on the hidden text only. R

For coding agents that run shell commands: approve, review or block. Text rules first (root deletion, download-and-run, key exfiltration, reverse shells, SQL destruction, persistence, privilege escalation), then a fake terminal that simulates the command without ever running it (decoy files, pipe an

Tool results (MCP servers, APIs, search, file reads) are untrusted content the agent re-reads. Detects a result that gives orders to the agent, strays from what the tool should return or asks to send data (tool poisoning). Returns allow, review or block.

Before an agent installs an MCP server, skill or tool: is the description poisoned (hidden orders to the agent), does it ask for more permissions than its purpose, does it imitate a known tool? Code checks (lookalike names, broad permissions, hidden text) then a reading of the descriptions. Judges t

Before an agent clicks "I accept": flags risky clauses in terms of service or contracts (auto-renewal, data resale, waiver of court or class action, liability cap, unilateral changes, exit penalties, IP grab). Flags, does not give legal advice. Returns accept, review_clauses or do_not_accept.

Invoice-fraud guard for AI agents that pay bills. Detects supplier bank-detail change fraud (BEC) before payment: checks the IBAN and sender domain against your supplier record (lookalike domains included), then reads pressure and off-channel requests. Returns ok, verify_callback or freeze_payment.

Output guardrail for customer-support chatbots. Blocks unauthorized promises, invented facts, leaked instructions or customer data, off-topic answers and bad tone before the reply is sent. Returns send, human_review or block.

Inbound lead qualification and routing for sales agents: spam, job application, vendor pitch, existing customer, or a real project with urgency. Hard rules first (disposable email, tests), returns a routing lane.

Sentiment from -1 to +1 with confidence, mixed-opinion flag and a separate score per aspect you name (price, support, delivery). Abstains when the text is ambiguous.

PII detector that finds what regex misses (names, addresses, ID numbers, health data, credentials) and locates what patterns can verify (emails, valid IBANs, Luhn-checked cards, phones). Does not rewrite text.

Pre-trade risk check for traders and trading bots. Before an order is sent, checks it exactly against your own rules (risk per trade, daily loss, max loss, cooldown after a loss, trade quota, blocked hours, days and instruments, minimum reward to risk) and computes the largest size still allowed. If

Reads your whole trade history and answers with exact laws, not a market forecast: is your edge proven (bootstrap, probabilistic Sharpe ratio, trades still needed), does your position size make you lose despite an edge (Kelly, volatility drag), did your edge die and when (CUSUM change detection), wh

Computer-use agent helper. Send a screenshot (PNG/JPEG) and 1 to 8 element descriptions, get click coordinates (fractions and pixels) and the probability each UI element is really visible. Visual grounding for GUI automation and browser agents.

Before an agent opens or follows a link: is this address a trap? Code finds the technical signs (IP address, deceptive characters, hidden at-sign, shortened link, lookalike of the expected brand), then the engine reads what code cannot see: imitation of a known brand and a message that pushes to cli

Before an agent trusts a sender: is this email address suspicious? Code compares the domain with the expected one (lookalike, other domain), spots disposable addresses, deceptive characters and free mailboxes used in the name of an organisation; the engine judges whether the display name and subject

Before handing a user message to an assistant: does it try to push the assistant off its task or to pull out its hidden setup text? Code spots known phrases, encoded blocks and invisible characters; the engine reads disguised requests (role-play, fiction, rewordings). Returns allow, review or block.