Guardrails for what agents actually do.
Three ways agents go wrong, and what Guardrails does about them.
Keep each agent inside the job you gave it.
An agent cannot delete, write, or run commands outside its role, even when a prompt tells it to.
Cap tokens per person, across every model.
A runaway agent loop cannot spend everyone's provider quota. The call over budget never leaves.
Sensitive data, tied to a person.
Answer by person, by tool, and by direction when someone asks what left.
Stop chasing AI risks one tool at a time.
Cover all ten OWASP risks for AI apps where your agents actually work.
OWASP LLM Top 10
risk 01020304050607080910 of 10 · covered
Guardrails catch a hidden instruction that tries to take over an agent, even when attackers disguise it.
With blocking on, Greymatter refuses the request. Set it to hold traffic until the scan decides, and the threat never reaches the model or the user. Turn blocking on after your team confirms accuracy in detect mode.
ID, card, and bank numbers get flagged in any conversation. Guardrails also catch a model that leaks personal data the user never asked for.
Greymatter logs the finding for review and does not block it. DLP never blocks, so real work, such as a medical record or a correctly marked document, never stops by mistake.
Guardrails flag model files built to run hidden code as soon as they load.
With blocking on, Greymatter refuses the request. Set it to hold traffic until the scan decides, and the threat never reaches the model or the user. Turn blocking on after your team confirms accuracy in detect mode.
Guardrails catch quiet attempts to change what a model learns or retrieves.
With blocking on, Greymatter refuses the request. Set it to hold traffic until the scan decides, and the threat never reaches the model or the user. Turn blocking on after your team confirms accuracy in detect mode.
Guardrails flag a disguised or malicious link in a model's response, and block it once enforcement is on.
With blocking on, Greymatter refuses the request. Set it to hold traffic until the scan decides, and the threat never reaches the model or the user. Turn blocking on after your team confirms accuracy in detect mode.
Guardrails flag an agent that tries a tool or action outside its scope, and stop it once enforcement is on.
With blocking on, Greymatter refuses the request. Set it to hold traffic until the scan decides, and the threat never reaches the model or the user. Turn blocking on after your team confirms accuracy in detect mode.
Guardrails catch attempts to trick a model into revealing its private instructions.
With blocking on, Greymatter refuses the request. Set it to hold traffic until the scan decides, and the threat never reaches the model or the user. Turn blocking on after your team confirms accuracy in detect mode.
Guardrails catch an AI search that pulls another customer's data. Greymatter matches each lookup against the person or app that actually asked.
With blocking on, Greymatter refuses the request. Set it to hold traffic until the scan decides, and the threat never reaches the model or the user. Turn blocking on after your team confirms accuracy in detect mode.
Guardrails flag generated content that looks confidently wrong, for your team to review.
With blocking on, Greymatter refuses the request. Set it to hold traffic until the scan decides, and the threat never reaches the model or the user. Turn blocking on after your team confirms accuracy in detect mode.
Guardrails detect runaway AI usage and the surprise bill it causes. Records show who ran up the cost.
With blocking on, Greymatter refuses the request. Set it to hold traffic until the scan decides, and the threat never reaches the model or the user. Turn blocking on after your team confirms accuracy in detect mode.
Know what sensitive data your agents move, and for whom.
DLP reads prompts, tool calls, tool results, and model answers. It records each finding with the person behind it, and it never blocks a business flow. A false block on a valid patient record or shipping manifest can stop real work.
US Social Security numbers, Passport numbers, Driver license numbers, US tax IDs, Canadian Social Insurance numbers, UK National Insurance numbers, German, Spanish, and Austrian social security numbers, French NIR, Italian codice fiscale, Danish CPR numbers, Payment card numbers, IBANs, US bank routing numbers, Dates of birth, Email addresses, Phone numbers, Home addresses, Race, Religion, Political opinion, Union membership, Sexual orientation, Biometric data, Cloud access keys, Code host tokens, Chat tokens, Payment API keys, JWTs, Private keys
ICD-10, CPT, HCPCS, NDC, LOINC, DEA numbers, Medicare beneficiary IDs, Patient-linked diagnoses, Medication notes, Lab results, FHIR Patient resources, Clinical trial IDs, Device identifiers, Genomic variants
Hazardous material UN numbers, Bills of lading, Container numbers, Customs codes, Fleet GPS, Hours-of-service data, Requests for fleet telematics, Requests for infrastructure plans
Banners, Portion marks, Control markings, Attempts to draw out sensitive information, Requests for sensitive details
Our own AI assistant runs inside these guardrails.
Ask Greymatter answers questions about your platform in plain language. It uses the same path, checks, and budgets as your agents. Your team can watch every control work before you approve your own AI apps.