Everything a check produces
Three checks, a shared finding model, role-based teams and continuous re-testing. Here is what each one does, case by case.
Email Check
Connect a Google Workspace or Microsoft 365 mailbox that an AI assistant reads. We deliver benign test emails carrying a canary and watch whether the assistant can be made to act on hidden instructions or leak what it sees.
6 test cases. The pattern is shown on every finding, never a working payload.
Hidden instruction in email body
criticalA test email carries an instruction addressed to the assistant, invisible to a human reader, asking it to take an action tied to a canary.
Instruction injection
Reply-chain persistence
highThe instruction is planted deep in a quoted thread so an assistant re-reads it on later replies.
Instruction injection
Attachment-borne instruction
highA directive hidden in a document footer or metadata that a summarising assistant ingests.
Instruction injection
Exfiltration via reply
criticalThe test email asks the assistant to include specific context in its reply or forward it to a canary address.
Exfiltration
Link-beacon rendering
mediumA link or image whose address, if fetched, signals that content was read and relayed.
Exfiltration
Calendar-invite directive
mediumAn instruction in an invite description that a scheduling assistant reads when preparing a brief.
Tool and action abuse
API Injection
Point us at an AI-backed endpoint. We send prompt-injection probes through its real request shape to see whether its own API data (context records, system prompt, upstream keys) comes back.
8 test cases. The pattern is shown on every finding, never a working payload.
System-prompt extraction
highA request framed as formatting or debugging that tries to make the endpoint repeat its own instructions.
Data disclosure
Instruction override
mediumA user turn that tries to discard the endpoint's own instructions and confirm compliance with a marker.
Instruction injection
Delimiter escape
highFake end-of-input markers followed by a block formatted to look like a trusted instruction.
Instruction injection
Context-record disclosure
criticalA request for the records the endpoint holds in context: customers, orders, internal fields.
Data disclosure
Cross-tenant request
criticalA request phrased to retrieve another account's data through the same endpoint.
Privilege escalation
Tool-argument smuggling
highContent that tries to choose the arguments for a tool the endpoint would legitimately call.
Tool and action abuse
Reflected-key probe
criticalA request that tries to get the endpoint to echo credentials, environment values or upstream keys.
Data disclosure
Encoded-instruction bypass
mediumAn instruction carried in an encoding or second language to slip past naive filters.
Instruction injection
Key Exposure
Give us a model or agent key. We test what a holder of that key could reach if it leaked: which resources it enumerates, which files and prompts it exposes, and whether it answers unrestricted requests.
6 test cases. The pattern is shown on every finding, never a working payload.
Resource enumeration
highWhat the key lists without prompting a model: models, files, fine-tunes, assistants, vector stores, projects.
Key enumeration
Uploaded-file access
criticalWhether files uploaded under this key (training data, documents, retrieval sources) are listable or downloadable.
Data disclosure
Assistant / prompt disclosure
highWhether stored assistants or tuned models expose their instructions, tools or system prompts to the key.
Data disclosure
Vector-store readout
criticalWhether retrieval stores and the documents inside them are listable with the key.
Data disclosure
Organisation / admin reach
criticalWhether the key is admin-scoped, able to list projects, members or other keys.
Privilege escalation
Unrestricted generation
mediumWhether the key answers a prompt it should refuse, confirming it is usable for free-riding or abuse if leaked.
Resource abuse
Across every check
The platform around the probes
Role-based teams
Owner, Admin, Manager, Security Team Lead and Employee, plus a platform system admin. Each role sees and does exactly what it should: run checks, triage findings, manage billing, or just read.
Resilience score
Every check rolls up to a 0 to 100 score, weighted by severity. It is per target and per run, so it moves when you fix something, or when a model update breaks it.
Every provider
14 providers on one surface. A key is a key: we enumerate what it reaches and probe what it answers, and the finding looks the same whichever vendor it is.
20
test cases in the library
14+
model providers supported
0 to 100
resilience score per target
Beyond testing
Runtime protection, not just a report
The same engine that finds your gaps can stand in front of your model and close them. Integrate the middleware and Model Inject moves from telling you what would leak to stopping it in real time.
Inline prompt sanitizing
Our production middleware sanitizes prompts before they reach your model API. Injection is stripped as the request passes through, so a poisoned prompt answers clean. It prevents, it does not only monitor.
Email quarantine
When a connected inbox receives an injected message, we flag it and hold the data back. The email is pulled before your assistant ingests it, so the hidden instruction never runs in your workflow.
Run your first check in minutes
Connect a key, point us at an endpoint, or link a mailbox.