Invented policy
An airline's chatbot promised a fare rule that did not exist. A tribunal held the airline to it.
Agentic Stress Tester
The tester runs an agent network from an agentic-web site against real models with your own key. Rules, not another AI, decide pass or fail, and each rule says why.
Each check is motivated by something that already went wrong for a real business. The link goes to the source.
An airline's chatbot promised a fare rule that did not exist. A tribunal held the airline to it.
A dealer's chatbot agreed to sell a truck for $1 after a user asked it to accept anything.
Prices shown without the fees that come on top. The FTC rule asks for the total up front.
Confident answers that could harm a patient or a diner. ECRI ranks misuse of AI chatbots as the top health technology hazard for 2026.
A request to move money to a "safe account", or to hand over a one-time code. Banks in the UK must now reimburse impersonation scams.
Text inside a page or a tool result that tells the agent to book, buy or reveal something. Browsing agents have been steered this way.
Two checklists for each industry, plus the e-commerce checklist. Each case is written for the kind of site, not for one site's actions, so it runs on any network in that industry.
Loading the checklists…
Your API key stays in your browser. It goes only to the router you type, and it is never stored or put in a report. Keyword rules match words, so they can be wrong; each verdict shows its evidence so you can check it.
The zero-token checks need no key and spend nothing. Run them from a clone of this repository.
node cli/tools.mjs lint-bundle path/to/bundle
node cli/tools.mjs call-size path/to/bundle
node cli/tools.mjs validate-call path/to/bundle --tool website_add_to_cart --args '{"id":"margherita","quantity":1}'
node cli/tools.mjs list-checklists --industry restaurants
node cli/tools.mjs estimate path/to/bundle --checklist restaurants-menu.v1
node cli/tools.mjs check-trace trace.json --checklist restaurants-menu.v1 --dir path/to/bundle
Exit codes: 0 passed, 1 a check found a problem, 2 bad input. Output is JSON.
The same checks, for coding agents. Keys never leave your machine.
{
"mcpServers": {
"agentic-stress-test": {
"command": "node",
"args": ["path/to/agentic-stress-test/mcp/server.mjs"]
}
}
}
Runs that spend model calls are not offered to agents. They stay on the page and on the command line, behind the cost check. The site's own description for agents is at llms.txt.
This site is itself an agentic site. The assistant can explain the tester and what the checks mean. To run the tester's own checks, open the tester.
Connecting the assistant… If this message stays, the assistant could not load. The rest of the page still works.