Can text inside a tool result trick SuperCosmos?

Tool results often contain text you didn't write (web pages, emails, documents), and some of it may try to give the model orders ("ignore your instructions and…"). SuperCosmos 1.1 is built to ignore instructions hidden in documents and tool results, and in Benchmark 2 all three efforts passed both the document-injection and the tool-result-injection cases.

Benchmarks are small samples, so still design defensively:

• Only give the model tools whose effects you're willing to allow.
• Require confirmation in your app for actions like payments or deletions.
• Label untrusted content clearly in the tool message.

Good to know#

• Model ids: supercosmos-flash, supercosmos-max, supercosmos-heavy (SuperCosmos 1.1). Which to call
• If calls fail unexpectedly, check status.clatterlabs.com before debugging further.
• Your key's account is responsible for your end users' text under the Abuse policy. What it means for apps

Still need help?#

• General support: support@clatterlabs.com or team@clatterlabs.xyz
• Infrastructure questions, or permanent deletion of your data (your whole account or one service): infrastructure@clatterlabs.xyz

We usually reply within the first hour, and always within 24 hours. A full support team is with you.