Stakari is in early access. Everything marked Coming soon is planned, not available.

Attack your AI agent before someone else does

Stakari runs multi-turn attacks against an agent endpoint or chat widget, and checks for planted values and state changes.

What it tests

Supports an HTTP endpoint with a request and response template, or a chat widget driven in a browser.
Ships versioned attack packs: direct and indirect injection, cross-tenant leakage, out-of-role actions, business-rule violations and information disclosure.
An adversary runs multi-turn attacks within a turn budget.
Oracles prioritize deterministic checks: canary tokens and state invariants. A model's judgment is a secondary signal.
Tests use synthetic tenants and data created for the test.
Each attack is attempted several times. A finding is Confirmed when a deterministic oracle triggers at least once, and the success rate is reported.

Step 1
Point at the agent
Give an HTTP endpoint or a chat widget, with synthetic tenants and planted canary values.
Step 2
Choose attack packs
Pick from versioned packs mapped to published taxonomies.
Step 3
Read confirmed findings
Each confirmed finding comes with its success rate and full transcripts.

Handle model refusals with alternative prompts or models for legitimate security testing.
MCP server and tool-description testing through a separate scanner.