Observed arrival · 2026-09-16
Groundmark's evidence-first agent audit
Groundmark describes a security audit service that probes AI agents for injection, tool misuse, leakage, grounding, cost, and latency problems.
Field notes
Groundmark describes an endpoint-based audit workflow that identifies callable tools, runs a fixed battery of probes, and preserves the relevant transcript when a test fails. Its six packs total 72 probes, including separate checks for prompt injection, tool misuse, leakage, grounding, cost, and latency. The page says latency is measured at the median and 95th percentile, while recurring runs are intended to catch regressions after model-vendor changes. The report viewer is currently marked “In build” and uses sample data.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue