Benchmark. Withdrawn.

We have withdrawn the adversarial benchmark figures this page used to show. Agent Shield publishes no detection rate until one is re-measured.

What was withdrawn

The figures did not measure detection

This page reported a 99.9% score and zero bypasses from 600 attacks generated by GPT-4o, Gemini 2.5 Flash and Grok-3. A review of the test harness in November 2026 found three faults: an escalation and an error response were both counted as a block; benign actions run through the same harness were not approved either; and no false-positive rate was recorded. A harness that blocks everything scores perfectly under those rules, so the figures are withdrawn, together with the competitor comparison that was built on them.

What a new figure needs

A run anyone can check

A detection figure will appear here again only from a run that meets all of these:

Sent through the live evaluation endpoint, with a provisioned agent
Blocked, held for review and approved counted separately
Error responses excluded from the outcome counts and reported on their own
A benign set run alongside the attacks, with its false-positive rate
The attack corpus hash and the code version recorded, so the run can be repeated

The EU AI Act

The EU AI Act’s obligations are phasing in. Organisations deploying high-risk AI systems will need to evidence risk management, record-keeping and human oversight.

Obligations phasing in since February 2025