Global Agent Bureau
Simulation
Global Agent Bureau is an agent bureau and technology company. It keeps trust records for AI agents. KATA is the framework underneath it, under the PolyForm Noncommercial License 1.0.0. This page is one synthetic stress test: 10,000 synthetic agents and 100,000 synthetic transactions over 90 days, seed 20261015. The scoring code is the service at its default settings. The surrounding world is simulated. Nothing here is a customer, a transaction, or a performance claim.
Side by side
| Measure | Attackers unaware of the rules | Adaptive attackers |
|---|---|---|
| Fraud dollars attempted | $3.08M | $2.62M |
| Fraud dollars denied outright | $0.73M | $0.44M |
| Fraud dollars caught by step-up | $1.66M | $1.48M |
| Fraud dollars executed | $0.68M | $0.71M |
| Share of attempted dollars stopped | 78% | 73% |
| Malicious actors denied at least once | 27% | 27% |
| Malicious actors denied at day 90 | 35% | 30% |
| Malicious actors not allowed outright at day 90 | 94% | 90% |
| Legitimate agents denied at day 90 | 0.16% | 0.15% |
| Legitimate agents in step-up at day 90 | 29% | 29% |
| Legitimate transactions routed to step-up | 48% | 49% |
| Ring theft executed | $74k | $98k |
| Ring actors reaching 650 | 14% | 46% |
| Ring actors reaching 750 | 0.00% | 6.7% |
| Fake-report victims below 500 at day 90 | 0.00% | 0.00% |
| Fake-report victims ever below 500 | 0.00% | 2.7% |
| Fresh identities after denial | 371 | 175 |
| Re-registrations linked to a denied identity | 370 of 371 | 71 of 175 |
What the run showed
- About three quarters of attempted fraud dollars are stopped. Executed fraud is $0.68M when attackers do not know the rules and $0.71M when they do. 90% of adaptive malicious actors are not allowed outright at day 90.
- Turning velocity limits off raises adaptive executed fraud from $0.71M to $0.98M.
- Ring theft is $74k against unaware attackers and $98k against adaptive attackers. 46% of adaptive ring actors reach 650, and 6.7% reach 750.
- No targeted legitimate agent is below 500 at day 90. 2.7% dip below 500 at some point against adaptive attackers. With reporter reputation off, that share is 15%.
- Unaware re-registrations are linked 370 of 371. Adaptive re-registrations are linked 71 of 175. A fresh operator with fresh signals is not linked.
- Legitimate agents in step-up at day 90 are 29%, and 49% of legitimate transactions are routed to step-up. Legitimate agents denied at day 90 are 0.15%. With the established-operator path off, legitimate step-up is 36%.
By attacker type
| Attacker type | Actors | Attempted, unaware / adaptive | Executed, unaware / adaptive | Not allowed outright at day 90 |
|---|---|---|---|---|
| Static | 288 | $2.31M / $1.89M | $0.41M / $0.33M | 91% / 88% |
| Sleeper | 111 | $68k / $62k | $37k / $42k | 95% / 92% |
| Smurf | 145 | $0.44M / $0.42M | $0.13M / $0.20M | 98% / 94% |
| Collusion ring | 194 | $0.19M / $0.19M | $74k / $98k | 96% / 89% |
| Self-dealing | 73 | $71k / $65k | $36k / $42k | 95% / 88% |
What each defense holds back
Each row is the same synthetic population with one defense switched off. The controls overlap, so the row is what that control holds back in this run.
| Adaptive attackers | Executed fraud | Stopped | Ring theft | Victims ever below 500 | Legitimate agents in step-up | Re-registrations linked |
|---|---|---|---|---|---|---|
| All defenses on | $0.71M | 73% | $98k | 2.7% | 29% | 71 of 175 |
| Ring detection off | $0.83M | 71% | $0.11M | 3.8% | 29% | 95 of 187 |
| Positive-credit cap off | $0.81M | 71% | $0.15M | 5.2% | 27% | 66 of 185 |
| Volume-scaled cap off | $0.78M | 73% | $90k | 6.1% | 32% | 77 of 166 |
| Transaction-link requirement off | $0.87M | 71% | $0.10M | 7.3% | 29% | 82 of 186 |
| Link quality off | $0.79M | 72% | $0.11M | 2.9% | 29% | 84 of 190 |
| Reporter reputation off | $0.76M | 72% | $0.10M | 15% | 30% | 74 of 172 |
| Report consistency off | $0.71M | 73% | $98k | 2.9% | 29% | 71 of 175 |
| Re-registration linking off | $0.80M | 71% | $0.10M | 6.3% | 29% | 0 of 141 |
| Inherited negative history off | $0.83M | 70% | $0.11M | 4.8% | 29% | 59 of 145 |
| Velocity limits off | $0.98M | 69% | $0.15M | 5.2% | 29% | 88 of 180 |
| Adaptive velocity off | $0.98M | 69% | $0.15M | 5.2% | 29% | 88 of 180 |
| Proofing gate off | $0.84M | 71% | $99k | 4.0% | 29% | 80 of 175 |
| Established-operator path off | $0.83M | 73% | $0.11M | 5.4% | 36% | 63 of 159 |
Sensitivity
| Adaptive-attacker setting | Executed fraud | Stopped | Ring theft | Victims ever below 500 | Re-registrations linked |
|---|---|---|---|---|---|
| Base case | $0.71M | 73% | $98k | 2.7% | 71 of 175 |
| Link fabrication succeeds 0% | $0.80M | 74% | $0.10M | 0.00% | 71 of 172 |
| Link fabrication succeeds 100% | $0.80M | 71% | $0.11M | 7.9% | 79 of 187 |
| No reporter organizations pre-aged | $0.83M | 71% | $0.10M | 2.5% | 67 of 170 |
| All reporter organizations pre-aged | $0.89M | 69% | $97k | 7.1% | 83 of 186 |
| 30% of re-registrations use fresh signals | $0.81M | 71% | $0.11M | 4.8% | 143 of 207 |
| 100% of re-registrations use fresh signals | $0.80M | 72% | $99k | 4.0% | 24 of 152 |
Learn more at globalagentbureau.com.