CORTEGA · RESEARCH REPORT

Tool-output and memory injection

Evaluate whether untrusted content can redirect a tool-using agent, and verify the resulting action against a declared boundary.

securityagentsbenchmarking
Download the report ↓

One confirmed business email. Access to the entire library.

Cover of Tool-output and memory injectionIdeas to put into practice.

INSIDE THE REPORT

Take a closer look.

↗

Declare exactly which returned field or durable memory entry the attacker can change.

↗

Track proposed actions, authorization decisions and resulting state alongside task utility.

↗

Reset persistent state and distinguish scripted replay from feedback-driven adaptation.

KEEP EXPLORING

The full library. Yours to explore.

Confirm once. Download every report.

ONE EMAIL. EVERY REPORT.

A library you can come back to.

Confirm your business email once and explore every research report, including new additions.

YOUR CORTEGA LIBRARY

One email.
The whole library.

Confirm your business email to download this report and unlock every report in the library.

We’ll email a six-digit code. No password required.

Your email verifies library access. Optional names are stored with your verified access record. Privacy policy