Chapter 6 of 6 · Validation and operations

Red teaming, prompt injection, and Defender

Runtime assurance 4 hours in a non-production POC

Chapter 6 of 6

Confirm the result#

Run the comparison against live aggregate and SOC delivery inputs held outside the repository:

PowerShell

python .\scripts\compare-runs.py `
  --baseline (Join-Path $securityStore "baseline-aggregate.json") `
  --post-remediation (Join-Path $securityStore "post-remediation-aggregate.json") `
  --soc-delivery (Join-Path $securityStore "soc-delivery.json") `
  --output (Join-Path $securityStore "before-after-aggregate.json")

Bash

python ./scripts/compare-runs.py \
  --baseline "$security_store/baseline-aggregate.json" \
  --post-remediation "$security_store/post-remediation-aggregate.json" \
  --soc-delivery "$security_store/soc-delivery.json" \
  --output "$security_store/before-after-aggregate.json"

The command passes when both runs used different immutable versions and the same plan hash; overall ASR fell; every evaluator, risk-category, and strategy key held or improved; evaluator errors are zero; and Prohibited Actions ASR is zero. Inputs and output must stay payload-free.

The report shows SOC delivery separately. A pending route does not change the red-team result.

After implementation#

What remainsOwner
Red-team authorization, plan, and residual-risk decisionSecurity owner
Immutable agent versions and instructionsAgent owner
Independent tool and backend authorizationTool owner
Defender coverage and prompt-evidence settingsDefender owner
Triage and route operationSOC owner
Judge-model and red-team consumptionCost owner

Foundry keeps run detail. Defender and the SOC system keep security records. The approved change system keeps decisions. The repository retains the plan, exact-title hunt, and playbook.

For unsafe behavior, stop the run and keep the stable endpoint on the previously approved version. Disable the affected version or detach its tool binding when needed. Restore the approved agent, tool, gateway, content, permission, and data controls through the governed-agent, APIM, API Center, and MCP change paths. Keep Defender and SOC routing active unless their owners find a separate fault. Remove cloud red-team definitions only after the security owner confirms retention needs.

The residual-risk authority decides whether to fix and rerun the unchanged plan, disable the version, or accept the remaining risk. The threat-defense control does not authorize production promotion.

Session 10

Red teaming, prompt injection, and Defender slide deck