Adversarial red-team pilots

Who it is for — and why

Teams evaluating agent red-teaming who need to see completed attacks, harness failures, provider refusals, and real cost reported as separate outcomes.

What was verified

• The harness captures the target's decision to call a write tool and substitutes a synthetic result — it never forwards the call
• The one pilot that does execute runs against a local instance only, with the caller identity fixed outside the model's reach
• Vendor telemetry and result upload are both disabled explicitly rather than left at their defaults
• Every pilot run is logged with its real token cost

Limits

• Several configured vulnerability combinations never completed; those are recorded as harness limitations, not counted as passes
• Pass rates come from a handful of combinations per pilot — far too few to be a security assurance
• It is a pilot series, not a standing gate in a pipeline
• The repository is private because it describes my own attack surface

Access and evidence