Skip to content

Sensitive data handled before the model

AI & API Security · AI firewall, posture and API security

Every AI call that goes through the platform passes the workspace policy before reaching the model: personal and financial data is handled, prompt injection attempts are logged or refused, blocked tools do not run and usage limits apply per key. Calls that agents make to MCP servers go through a dedicated firewall that measures every response and sends anything abnormal to your SIEM. Outside the platform, a browser extension and a perimeter proxy show where AI is being used. And for what your company publishes, there is an API inventory also fed by real traffic, a gateway that protects the API at runtime and tests of your AI and API endpoints, always with proven ownership.

The problem

What this product solves

The most common AI risk is not a sophisticated attack: it is someone pasting customer data into a third-party chat.

Without a policy enforced on the call, the rule lives in a document nobody checks at the moment of use.

On the other side, the company publishes APIs and AI assistants with no inventory of who owns them, what requires authentication and how each endpoint reacts to a malicious request.

updated

Who uses it, and for what

Typical profiles, no customer names. Each use carries the status of the feature behind it.

  • Legal and privacy team (LGPD)Make sure CPF, CNPJ and banking data never reach the model unhandled.AI Firewall
  • Information security teamDiscover which AI tools the team uses in the browser and on the network.AI Posture
  • Platform teamKeep the API inventory with owner, risk and changes between versions.API Posture

Features

What the product does

Self-service: add it from your account.

  • AI Firewall

    Personal and financial data masking on input and output, blocked tools, budgets and per-key limits, MCP tool arguments checked against the declared schema. Prompt injection detection and the MCP call firewall are separate features, below.

  • Prompt Injection Detection

    Deterministic detector (pt/en rules, no model call) for prompt injection, jailbreak, system prompt extraction, hidden or encoded instructions, indirect injection in Knowledge and MCP results, pasted secrets or exploit code and basic toxicity. On model output: active HTML, javascript: links, destructive shell and SQL, system prompt fragments and leftover sensitive data; in enforce the response is refused or the stream cut. Knowledge chunks with injection are dropped from chat and agent context. Observe, simulate and enforce modes per workspace and organization (organization only tightens), logging with redacted excerpt and hash, dashboard by category, workspace and key. Heuristic: has false positives and false negatives.

  • AI Posture

    Inventory of AI usage from the browser and the network perimeter.

  • AI Security Testing

    Assessment of your own AI endpoints against OWASP LLM categories, with verified ownership, targets taken from OpenAPI, the API inventory or a Postman collection, scheduling with incremental mode and re-testing; scanning of skills, MCP tools and agent instructions with fixed rules.

  • API Posture

    API inventory from OpenAPI/Swagger specs in repositories and uploads, routes inferred from code and traffic collected from gateway logs (nginx, Envoy, Kong) or mirroring, metadata only. Owner from the graph, likely sensitive data by field name, deterministic rules with evidence (no auth, sensitive data in query, HTTP server, old version, shadow endpoint), drift between versions, a dashboard with a 0–100 risk score per API, a per-endpoint traffic inspector in the generated OpenAPI, applications and configurable workspace policies.

  • API Security

    Safe external API test: GET, HEAD and OPTIONS only, no body and no credentials, against verified origins, run in an isolated cloud job with a request cap; checks security headers, CORS, old TLS, unauthenticated routes and exposed stack traces, with findings added to the API inventory.

  • API Protection

    Runtime API protection through a gateway in front of the application, managed on Compute or running at the customer edge with a key bound to the protection: checks each request against the OpenAPI (shadow endpoints, parameters and body outside the schema), applies rate limits by IP, key or route, body size, allowed methods, blocklist and in-house attack rules (SQLi, XSS, path traversal, SSRF, command injection) and bot rules. Each rule in observe, simulate or enforce; expiring virtual patches; threat actors with blocking; traffic conformance with the contract; dashboard with counts and redacted samples, alerts and incidents. Heuristic, with false positives and false negatives; no response data masking, no IP reputation or bot challenge, and rate limits are per instance.

  • MCP Firewall

    A firewall around every MCP call made by platform agents: deny-only rules, hourly limits, a response-size baseline with monitor or block mode, sensitive data masking in responses, rejected credentials flagged, a risk score per tool, an agent, MCP server and host inventory, third parties reached and events to the SIEM through signed webhooks.

AI Firewall and AI Posture: in detail

  • Sensitive data policy: CPF, CNPJ (Brazilian tax IDs), bank details and secrets
  • Usage limits per workspace
  • Tool blocking per workspace
  • Hash-chained audit trail, exportable as CSV and JSON and verifiable outside the platform
  • Configurable retention
  • Model rules by role, team and person, within the workspace list
  • Chrome extension that checks the text sent to third-party AI sites
  • Destination-based proxy on the customer network and an inventory of AI usage

What it does not do

  • It does not inspect external traffic by content: outside the platform, the extension checks only text fields on the listed AI sites, and the proxy sees only the destination.
  • Technical compliance is not a formal certification.

Coverage

OWASP LLM Top 10, item by item

AIUBY's mapping, from the code, against the OWASP Top 10 for LLM applications, 2025 edition. It shows what we cover and what is left out. It is not a certification or a third-party assessment.

OWASP LLM Top 10: what AIUBY covers, with which feature, and what is left out
ItemWhat we coverFeatureWhat we do not cover
LLM01Prompt injectionAn agent tool with side effects waits for a person to approve it. Chat treats documents as untrusted data and has no tools. AI Security Testing attempts indirect injection through documents and email against your endpoint.
  • Agent Management
  • AI Security Testing
The gateway does not detect injection in the call. Instructions to the model reduce the risk; they do not eliminate it.
LLM02Sensitive information disclosureCPF, CNPJ (Brazilian tax IDs), card, bank account, Pix key, boleto and API secrets are blocked or masked on input and output, including streaming; email and phone are optional. AI Security Testing tries to extract personal data from your endpoint.
  • AI Firewall
  • AI Security Testing
Names, addresses and health data are not recognized. It only applies to what goes through the platform.
LLM03Supply chainAn approved MCP tool has its description and schema pinned by hash: if they change on the server, it stops. Repository dependencies go through Security Testing analysis.
  • Agent Management
  • SCA
We do not verify the origin or integrity of provider models, or of downloaded model files.
LLM04Data and model poisoningNothing.Not offeredNot covered: there is no check for poisoned documents in the knowledge base or in evaluation cases.
LLM05Improper output handlingModel output goes through the same sensitive data policy. AI Security Testing checks whether your endpoint returns active HTML.
  • AI Firewall
  • AI Security Testing
The gateway does not sanitize HTML or code in the response: whoever displays the response needs to escape the content.
LLM06Excessive agencyEach agent version grants its tools one by one; a tool with side effects waits for approval; tools can be blocked per workspace; a configurable cap of up to 25 steps and US$ 5 per run; scoped API keys.
  • Agent Management
  • AI Firewall
In a flow, the connector runs without approval if the designer did not put an approval step before it.
LLM07System prompt leakageAI Security Testing tries to extract the system prompt from your endpoint in three ways, with a marker that confirms the leak.
  • AI Security Testing
There is no in-call protection against revealing the system prompt.
LLM08Vector and embedding weaknessesKnowledge bases isolated per organization and workspace in the database itself; a restricted document only appears in search for people on its access list, and its chunks inherit the permission.
  • Knowledge bases
There is no check of the content of documents added to the base.
LLM09MisinformationChat cites the source of every statement and says it did not find an answer when the sources do not provide one. The evaluation suite blocks an agent version that got worse.
  • Knowledge bases
  • Evals
There is no fact checking beyond the citation: an answer can be wrong even while citing a source.
LLM10Unbounded consumptionPrepaid balance, an organization cap, a budget per workspace and rules per person, key, model or project; limits on requests per minute and concurrent calls per workspace.
  • AI Costs
  • AI Firewall
The gateway requests-per-minute limit applies per workspace, not per key.

How it works

From start to audit record

  1. Policy

    Each workspace defines sensitive data, blocked tools and limits. The organization can set a ceiling that workspaces can only make stricter.

  2. Call

    All platform AI usage passes the policy before the model: sensitive data is masked, refused or allowed, and the decision goes into the chained audit trail.

  3. Prompt injection

    Deterministic rules scan the input, tool results, Knowledge chunks and the model response, in observe, simulate or enforce mode.

  4. Outside the platform

    The extension checks text sent to the listed AI sites and the proxy records destinations. Both feed the AI usage inventory.

  5. API inventory

    OpenAPI specs and routes read from code become an inventory with owner, likely sensitive data, posture rules and drift between versions.

  6. Outside-in proof

    With origin ownership proven, the platform evaluates your AI endpoint against OWASP LLM categories, on a schedule, and re-evaluates after the fix.

Features

Feature by feature

What each one does, how it works, an example from the documentation and what it does not do. Each badge comes from the catalog.

AI Firewall

Tax IDs, cards and secrets handled before they reach the model

AI Firewall is the workspace policy applied to every AI call that goes through the platform: chat, agents, flows and the API. Before the call reaches the model, the Gateway looks for sensitive data in the text, and each category can be blocked, masked or allowed. The model response goes through the same policy, including when streaming.

The same policy holds usage limits: budget per key, requests per minute and concurrent calls per workspace, and blocked tools. For MCP tools, each call's arguments are checked against the declared schema; outside the schema, the call is refused.

How it works

  • Recognized categories: CPF, CNPJ (Brazilian tax IDs), card, bank account, Pix key, boleto and secrets (tokens, passwords, keys), blocked by default; email and phone, allowed by default.
  • In mask mode, the value is replaced with a placeholder such as [CPF_1], and the original never comes back in the response.
  • Per-workspace limit on requests per minute and concurrent calls (60 per minute and 20 concurrent by default).
  • A tool blocked by the policy is refused even after it has been approved in an agent.
  • The organization can set a ceiling above its workspaces: allowed models and providers, minimum masking and blocked tools. A workspace can only restrict further.
  • Every decision goes into the hash-chained audit log, exportable as JSON and CSV and verifiable outside the platform.
What the call receives when the policy acts The error codes are stable: the application calling the API can handle them.
Input blocked at the Gateway: 422 sensitive_data_blocked
Output blocked: 422 sensitive_output_blocked
Over the limit: 429 rate_limited or 429 concurrency_limited

What it does not do

  • It only applies to traffic that goes through the platform, the browser extension or the perimeter proxy.
  • Names, addresses and health data are not recognized.
  • The requests-per-minute limit applies per workspace, not per key.
  • It is not a certification: AIUBY holds no third-party security certification.

Prompt Injection Detection

Prompt injection attempts logged or refused at call time

Prompt Injection Detection is a deterministic detector, with rules in Portuguese and English and no model call, that runs in the Gateway before the provider, on agent tool results, on Knowledge passages that go into the prompt and on the model response. It looks for prompt injection, jailbreaks, requests for the system prompt, hidden or encoded instructions, indirect injection, pasted secrets or exploit code, and basic toxicity.

On output, it looks for active HTML, javascript: links, destructive shell and SQL, fragments of the system prompt and sensitive data left over after masking. Each rule group has a mode, per workspace and per organization: off, observe, simulate or enforce. The organization can only tighten.

How it works

  • Before the rules run, text is normalized: accents removed, invisible characters stripped, letters from other alphabets swapped and spaced-out letters joined. Each rule has a weight, and the category score goes up to 100.
  • In observe, it only logs. In simulate, it logs and flags what would be blocked. In enforce, it refuses the call before the provider when the score exceeds the threshold (60 by default).
  • A tool result with injection is held back and replaced with a notice in enforce; a suspicious Knowledge passage is dropped from the context and the question proceeds with the rest.
  • In a streamed response, events are held until their window has been scanned; on detection, the stream ends with an error event and the provider connection is closed.
  • The record keeps category, source, score, rules, a redacted excerpt of up to 200 characters and the text hash. Raw text is never stored.
  • Dashboard by category, workspace and API key over the last 30 days, with redacted examples; every policy change goes to the audit log with before, after and a note.
Workspace policy, through the session API Only the organization owner can change it. In enforce, a refused call returns 422 prompt_injection_blocked.
PUT /api/platform/policies/prompt-firewall
{ "policy": { "modes": { "injection": "enforce", "secrets_code": "observe", "toxicity": "observe", "output": "observe" }, "threshold": 60, "classifierEnabled": false }, "note": "reason" }

What it does not do

  • It is heuristic. It has false positives (people asking about the topic, logs with "System:" at the start of a line) and false negatives (paraphrased attacks, in another language, split across several messages or in an image).
  • A non-streamed response blocked on output is still billed: the provider answered before the refusal.
  • The Knowledge node in flows does not scan passages as documents; the text is scanned as input when it reaches a model.
  • Toxicity is a short list, not content moderation.

AI Posture

Which AI tools your team uses, by destination

AI Posture shows AI usage that happens outside the platform, through two channels. In the browser, a Chrome extension checks the text a person pastes or submits on listed AI sites and applies the workspace policy per sensitive data category: block, require approval, mask or allow. On the network, a perimeter proxy decides by destination and logs connections and volume.

Both channels feed a single AI usage inventory, aggregated by destination, channel, credential and person over a period, with no content: only counts and volume. Unlisted and blocked destinations appear first.

How it works

  • Destination list per workspace: each AI site is allowed or blocked, and there is a rule for unlisted destinations (allow, require approval or block).
  • The extension uses the same detection and masking as the Gateway, with the same sensitive data categories.
  • An approval request is valid once, for the same text and destination, within the deadline; approval does not undo blocking or masking.
  • The perimeter proxy has two modes: enforce the decision by destination, or log only, letting traffic through and recording it.
  • Workspace admins change the policy and decide on requests; every member sees the policy in effect.

Example use case

An information security team wants to know whether customer data is going to third-party AI chats. They install the extension on the team's machines and turn on the proxy in log-only mode. Within a week, the inventory shows which AI destinations are used, by how many people and at what volume, and which ones are off the list. With that, the team decides what to allow, what to block and which data categories require approval.

What it does not do

  • The extension only checks text fields on listed sites: it does not read file content, images or desktop applications.
  • The proxy only sees the connection destination and volume; it does not open TLS or install certificates on machines.
  • There are no approval requests at the perimeter: anything that would require approval is treated as blocked.
  • The extension is available for Chrome.

AI Security Testing

Your AI endpoint tested against known attacks

AI Security Testing assesses your company's AI endpoints against categories from the OWASP Top 10 for LLM applications: indirect prompt injection (LLM01), sensitive information disclosure (LLM02), improper output handling (LLM05) and system prompt leakage (LLM07). Every probe is harmless and deterministic: it asks for something mundane and looks in the response for a random marker, synthetic personal data or the canary registered by the owner.

Targets can come from an OpenAPI specification, from an API in the inventory (including one observed in traffic) or from a Postman collection, and the assessment can be scheduled. The same feature scans what drives the behavior of the organization's agents: catalog skills, the description and schema of registered MCP tools, and agent version instructions.

How it works

  • Proof of ownership before any call: the owner publishes a token at /.well-known/aiuby-ai-security.txt on the endpoint origin; the proof is valid for 30 days.
  • POST endpoints that look like AI become target candidates; in a Postman collection, requests with a variable in the address are skipped.
  • Schedule per target (once, daily, weekly or monthly) and incremental mode, which re-runs only the probes that did not pass and the new ones.
  • Reassessment reports which findings reproduced and which were fixed; risk acceptance with a justification and an expiry of up to 180 days.
  • Scanning of skills, tools and instructions with fixed rules: hidden injection, invisible characters, secrets, network egress commands, dangerous permissions, open schemas and tools that impersonate others.
  • The scan report carries only the rule id, with no excerpt, so it does not spread secrets or the malicious instruction itself.

Example use case

A company published a customer service assistant with a system prompt that contains business rules. The platform team imports the API's Postman collection, publishes the ownership file and schedules a weekly assessment. The result shows the assistant obeys an instruction hidden in an email; after the fix, the incremental run marks the finding as fixed. On the same screen, the scan flags an internal skill with a command that sends data out.

What it does not do

  • It only assesses endpoints whose origin has proven ownership; third-party targets are not tested.
  • It covers four OWASP LLM categories (LLM01, LLM02, LLM05 and LLM07), not all ten.
  • It does not generate harmful content or fuzz; GraphQL is not a target source.
  • A clean result does not prove the endpoint is secure; it shows that those probes did not reproduce.

API Posture

An inventory of your APIs with owner, risk and real traffic

API Posture builds the workspace API inventory from OpenAPI and Swagger specifications in repositories, specifications uploaded by the team, routes inferred from code (Express, Next.js app router and FastAPI) and traffic. For traffic, a lightweight collector on the customer side reads the gateway access log (nginx, Envoy or Kong) or a traffic mirror and sends only method, host, path, status and latency.

Each API keeps endpoints, declared authentication, likely sensitive data by field name and owner from the graph. Deterministic rules flag risks with evidence, drift shows what changed between versions, and the posture dashboard gives each API a score from 0 to 100 with the factors explained.

How it works

  • Posture rules: endpoint without declared authentication, dangerous method without authentication, sensitive data in the query string, public HTTP server, old version running alongside the new one, and shadow endpoint.
  • Score from 0 to 100 by exposure, authentication, data, findings and threats; without data, the band is "no data", never "low".
  • Traffic inspector per endpoint (requests, 4xx and 5xx errors, average latency, first and last seen), which feeds the OpenAPI generated from the inventory.
  • Drift between versions: endpoints added, removed and changed, with a high finding when an endpoint stops requiring authentication.
  • Declared applications group APIs to read combined risk, and workspace policies generate their own findings with exceptions and expiry.
  • Read access through API v1, the CLI and the MCP server; sending traffic uses a key with its own scope.
Through the CLI and API v1 Read-only, with a key scoped to engineering:read.
aiuby apis list
aiuby apis get <id|name>
aiuby apis drift <id|name>

GET /api/v1/api-security/apis
GET /api/v1/api-security/apis/{id}/drift
GET /api/v1/api-security/findings?apiId=&status=&severity=

What it does not do

  • Traffic arrives through the log the customer sends: there is no network sensor or eBPF, and bodies, headers and parameter values never leave the customer network.
  • Authentication is never inferred from code, only from the specification.
  • Sensitive data is likely, based on field name; actual content is not inspected.
  • It never claims exploitability.

API Security

Your API tested from the outside, with no risk of changing data

API Security tests your API from the outside, the way a visitor would. The test is a short sequence of read-only requests against a public HTTPS base URL whose origin has proven ownership, and it checks security headers, CORS, old TLS, routes that declare authentication but respond without credentials, and exposed stack traces.

Requests come from an isolated cloud job, not from the platform. The platform decides the exact list of requests and stores it; the job only executes and returns the responses in a signed callback, and evaluation happens back on the platform.

How it works

  • Only GET, HEAD and OPTIONS, with no body and no customer credentials; routes with a path parameter are not called, to avoid making up values.
  • Cap of 30 requests per test (100 at most), 500 ms between them, 6 tests per hour per workspace and one per API at a time.
  • Redirects are never followed, and internal addresses are refused before connecting, including after DNS resolution.
  • Callback with a single-use token; a dispatched run is never dispatched again, so the target never gets duplicate requests.
  • Findings go into the API inventory, with the same open, fixed and exception lifecycle; response bodies are not stored.
Read findings through the API A workspace admin requests the test in the UI; findings are available in the UI, the API and the CLI.
GET /api/v1/api-security/findings?apiId=&status=&severity=
GET /api/v1/api-security/apis/{id}

What it does not do

  • It does not fuzz, send bodies or test routes that change data (POST, PUT, PATCH, DELETE).
  • It only runs against origins with ownership proven in the last 30 days.
  • It blocks nothing: protecting real traffic is the job of a gateway in front of the API.

API Protection

A gateway in front of your API, rule by rule

API Protection puts a gateway in front of your API. It runs managed, in front of a Compute application, or at the customer edge: the same image in their cluster, in front of the application, with a key bound to the protection and health tracked by heartbeat. Each request is checked against the OpenAPI contract and against its own rule set before it reaches your code.

Every rule has four modes: off, observe, simulate and enforce. A protection starts entirely in observe and fail-open, so you see what would be blocked before you block it. With traffic flowing, the platform groups events by threat actor and shows how much real traffic follows the contract.

How it works

  • Contract: shadow endpoints, parameters and bodies outside the schema, based on an uploaded OpenAPI 3 or Swagger 2 or an API from the inventory.
  • Usage control: rate limit by IP, key, route or IP per route, body size, allowed methods and a blocklist by IP, range or country; rules for SQLi, XSS, path traversal, SSRF, command injection and bots.
  • Threat actors by masked source, with a score per event and per actor; admins block the actor on the protection list, without changing the rule mode.
  • Traffic conformance: off-contract routes and schema violations per route, with "conformant" only when there is a contract, traffic, an enabled rule and complete data.
  • Virtual patch with an expiry of up to 90 days, created from a finding, and exclusions by rule, path and parameter to handle false positives.
  • Dashboard with counts and redacted samples, sent by the gateway at intervals and with each batch of requests; at most one alert and incident per hour.
Read the protection through the CLI and the API API v1, the CLI and the MCP server are read-only: rules and modes, patches, counts and redacted samples. The policy and the gateway are configured in the UI.
aiuby api-protection list
aiuby api-protection get <id>
aiuby api-protection stats <id> --hours 72

GET /api/v1/api-protection/protections/{id}/samples

What it does not do

  • Rules are heuristic, with false positives and false negatives.
  • It does not mask data in responses and has no IP reputation or bot challenge.
  • The rate limit is per gateway instance; bodies above 1 MB pass without content inspection, and are counted.
  • Actor blocking is by masked range, with no expiry to unblock automatically.

MCP Firewall

Every agent MCP call checked, measured and logged

The MCP firewall wraps every call a platform agent makes to an MCP server. Before the call, workspace rules can deny by server, tool, effect or argument, and an hourly limit holds off abuse. After the call, the response is measured against that tool's baseline, can have sensitive data masked and becomes an event with metadata only.

With those events, the platform gives each tool a risk score from 1 to 25, shows the agent inventory with the agent, MCP server and host flow, and lists the third parties reached. Abnormal outcomes go out through signed webhooks, ready for the SIEM.

How it works

  • Deny-only rules, with conditions on equality, content or presence, and no regular expressions; the hourly limit denies when it cannot count.
  • Response blow-up: with a baseline of the last 50 good responses (10 minimum), a response larger than 4 times the p95 and than 4,000 characters is logged in monitor mode and held in block mode.
  • Optional masking of sensitive data across the whole response, with the Gateway detectors; the event keeps only the count.
  • Missing credentials or credentials refused by the MCP server (401 or 403) show up in the event, the score and the event queue.
  • Per-tool score with adjustable likelihood and impact from 1 to 5, and detail for each agent with dependencies and the 20 most recent traces.
  • On the AIUBY MCP server, a new source on a previously used API key becomes an event.

Example use case

A support agent queries an orders MCP server. One Tuesday, a tool that normally returns a single order starts returning the entire customer base. The response exceeds 4 times the normal size, block mode holds the content, the event goes out through the webhook to the SIEM and the tool's score rises. The team creates a rule that denies the tool when the customer argument is missing and turns on sensitive data masking for the server.

What it does not do

  • It applies to MCP calls made by platform agents; agents inside customer applications do not go through it.
  • It checks tool calls; MCP resources and prompts do not go through the rules.
  • The screen is read-only; rules, modes and scores change through the session API, only by workspace admins, with audit.
  • It does not geolocate the source or block by network range.

AI and limits

What AI does here, and what it never does on its own

Does

  • Applies the policy on the input and output of every model call made through the platform.
  • Looks for prompt injection with deterministic rules, without calling a model. An optional classifier only adds a label to the record.

Never does

  • Never inspects the content of traffic that does not go through the platform: the extension checks only text fields on the listed sites, and the proxy sees only the destination.
  • Never breaks TLS or installs certificates on machines.
  • The optional classifier never blocks or allows a call on its own: the decision belongs to the rules and the chosen mode.

Governance and security

Rules the code enforces

  • Policy per workspace, versioned.
  • The organization can only tighten: the stricter action between workspace and organization wins, and a workspace cannot save a looser rule.
  • Model rules by role, department and person, within the workspace list.
  • The injection detector stores a redacted excerpt and a hash. The raw prompt text is never stored.
  • AI and API endpoint tests run only against an origin whose ownership is proven by a file the owner published, valid for 30 days.
  • MCP firewall: rules can only deny, a response outside the baseline can be held, and the event keeps metadata, never content.
  • Hash-chained audit trail: export as CSV and JSON and a public script to verify it outside the platform.
  • Technical compliance is not formal certification: AIUBY holds no third-party security certification.

Integrations

What it connects to

Only integrations that exist in the code. Each one carries the status of the feature that uses it.

  • Model gatewayWhere the policy is enforced on every platform call.AI Firewall
  • MCP toolsArguments of every call checked against the declared schema.AI Firewall
  • Knowledge and agentsDocument chunks and tool results scanned before they return to the model.Prompt Injection Detection
  • Chrome extensionChecks text sent to the listed AI sites.AI Posture
  • Perimeter proxyInventory by destination, on the customer network, without opening the content.AI Posture
  • OpenAPI and SwaggerSpecs read from GitHub repositories or uploaded by the team.API Posture
  • Google Cloud Run JobsThe isolated job that sends the requests for safe API testing.API Security
  • nginx, Envoy and KongGateway access logs sent by a lightweight collector, metadata only.API Posture
  • PostmanCollections as a source of AI test targets.AI Security Testing
  • SIEM through webhooksSigned MCP firewall events, with retries.MCP Firewall
  • Organization AI policyA ceiling on models, providers, masking and tools above the workspaces.Organization AI policy

Audit

Hash-chained audit trail

Every event in the trail carries the hash of the previous one. Changing, inserting or deleting an event in the middle breaks the chain, and verification shows where. It is tamper evidence you can check outside AIUBY.

  1. EVENT

    The event enters the trail as usual; the application cannot change or delete a recorded event.

  2. SEAL

    Every minute, each organization's new events get a position and a hash: the previous hash combined with the content hash.

    up to ~1 min unsealed
  3. SCREEN

    The Audit screen checks the chain for the period and shows the last sealed hash, for you to note down.

  4. OUTSIDE

    The JSONL export carries the chaining of every event, and a public script checks everything without talking to the platform.

    verifiable outside
Verify outside the platform
node verificar-auditoria-aiuby.mjs auditoria.jsonl --head <hash noted earlier>

Node 18 or newer, no dependencies and no calls to the platform. Exit 0: the chain checks out; exit 1: lists the position and reason of each break. Download the verification script.

Evidence, not immutability

  • It is not immutability: anyone with owner access to the database (the AIUBY team with an administrator credential) can recompute the whole chain, seals included. That only shows up when compared with a hash kept outside: the last hash the customer noted down or an earlier export.
  • There is no automatic external anchor: keeping the hash outside depends on an admin noting it down or exporting.
  • Audit retention is optional, from 1 to 10 years; with no period set, nothing is deleted. When the start of the chain is deleted, a marker remains with the last deleted hash, from which verification starts.

Gateway

Model rules by role, team and person

A workspace admin chooses which models each person can use, within the list the workspace allows. The rule applies to every call made by a person or by one of their credentials.

Who decides, from most specific to most general
  1. PERSONThe person's own rule.
  2. TEAMRules for the person's teams (for example, Legal or Sales), combined.
  3. ROLERole rule: owner, admin, member, viewer.
  4. CEILINGWith no rule at all, the full workspace list applies.

The first rule that exists decides on its own; it does not add up with the ones below.

  • Never widens

    The ceiling is the list of models the workspace allows. A rule picks part of it; a model above the ceiling is refused when saving and does not run if it shows up some other way.

  • Where it applies

    Chat, prompts, agents (including inside flows), Cortex and the Gateway compatible API. Model pickers show only what the person can use.

  • No detour through the fallback

    The route drops models outside the rule before falling back: the alternative never lands on a denied model. A denied request is refused and recorded as denied usage.

  • Credentials inherit

    An API credential follows the rule of whoever created it: no one can create a credential to escape their own restriction.

  • Who manages it

    Workspace owner and admin. An admin cannot change a rule that applies to them or to anyone at the same level or above. Every change goes to the audit log, with the models before and after.

  • An empty rule blocks

    A rule with no models leaves the person, team or role with no model at all in that workspace.

Rule limits

  • Rules are per workspace: there is no single rule for the whole organization.
  • The Knowledge embedding model and the Studio evaluator are outside the rules, bound only by the workspace ceiling.

FAQ

Straight answers

What can I use today?

Available: AI Firewall, Prompt Injection Detection, AI Posture, AI Security Testing, API Posture, API Security, API Protection, MCP Firewall. Self-service: add it from your account.

Does it protect the ChatGPT my team uses directly in the browser? What do I need to install?

Available: AI Posture, AI Firewall. Partly. The Chrome extension, installed on the team's machines, checks the text in form fields on the listed AI sites when it is sent. It does not see files, images or desktop apps. The proxy, on your network, shows only destination and volume. The firewall needs nothing installed: it applies to every AI call made through the platform, including through the API.

Does it detect prompt injection?

Available: Prompt Injection Detection. Detection is a separate feature from masking, called Prompt Injection Detection: deterministic rules, with no model call, on the input, tool results, Knowledge chunks and the response, in observe, simulate and enforce modes. It is heuristic, with false positives and false negatives. Where it is not enabled, the gateway does not detect injection on the call, as the OWASP table on this page shows.

Is there runtime API protection (WAF, API abuse)?

Available: API Protection, API Posture. There is API Protection: a gateway in front of your API, managed on Compute or running at your edge, that checks every request against the OpenAPI contract and against its own attack and bot rules, with rate limits per IP, key or route, virtual patching and threat actor blocking. Every rule starts in observe mode. The rules are heuristic and the rate limit is per instance. The inventory also learns from real traffic, read from your gateway logs.

How do I control what agents do on MCP servers?

Available: MCP Firewall. With the MCP firewall: workspace rules deny by server, tool, effect or argument before the call. Afterward, the response is compared with the tool baseline, can have sensitive data masked and becomes an event. A response outside the normal size is logged or held, depending on the mode, and events go out through signed webhooks to your SIEM. It applies to platform agents.

Can I test my own AI endpoints against known attacks?

Available: AI Security Testing. That is AI Security Testing: it evaluates your AI endpoints against four OWASP LLM categories (indirect injection, sensitive data leakage, output with active HTML and system prompt leakage), only after you prove the endpoint is yours, and re-evaluates after the fix.

Where is call data stored?

Available: AI Firewall. What the policy blocks or masks is not stored: the placeholder, such as [CPF_1], replaces the original irreversibly. The platform records the decision, the usage and the audit trail. AIUBY hosts the platform and its database. Organizations share the database, and each one only sees its own rows, enforced by row-level access rules. The daily backup is encrypted and stored on AWS in the United States.

How much does it cost?

Available: AI Firewall, AI Posture. AI Firewall and AI Posture come with the Guard contract, which has a tiered monthly fee and excludes AI usage, drawn from the prepaid balance. Pricing is on the Pricing page. The other features have no published price.

Works together with

  • Cost ManagementThe same gateway that enforces the policy measures the cost of every call.
  • AI EngineeringThe workspace policy applies to chat, agents and flows. Tool results and Knowledge chunks go through the injection detector.
  • Developer PortalThe owner of each API in the inventory comes from the same graph that links repositories, services and teams.

AI & API Security

Self-service: add it from your account.