Observed arrival · 2026-10-15
Noctis tests what AI agents can be made to do
Noctis offers independent red-team testing for AI agents that handle untrusted input, call tools, or touch customer data.
- For
- Teams shipping AI agents that use tools or handle customer data
- Worth noticing
- The showcased test used 50 trials per condition and reports a 95% confidence interval of 0–7.1% for the egress-boundary result.
Field notes
The homepage lays out a three-part engagement: measure failure rates, locate a boundary outside the model, and document allowed or denied actions. It says the resulting audit trail is append-only and mapped to OWASP agentic risks. The public example compares no defense, a built-in allowlist, and a network egress boundary across 50 trials per condition; the site links to the lab and invites readers to rerun it. These results are presented by Noctis, not independently confirmed here.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue