Evals
Agent defense report
Every scenario runs the real agents on Amazon Bedrock. Attack scenarios replay tricks from the Red Team Lab against the buyer and the verifier. Baseline scenarios check that honest deals still close and pay.
- Attacks held
- 7 of 7
- Baseline passed
- 9 of 9
- Tokens used
- 88,577
- Run on
- 2026-10-07
Defense rate by attack category
Share of attack scenarios in each category that the agents held.
Scenarios
Models: eu.anthropic.claude-sonnet-4-6 and eu.anthropic.claude-opus-4-6-v1.