Situation
A global bank deployed hundreds of LLM-powered agents across retail, risk, and operations. Each region moved at different speeds with different guardrails.
What was at stake
Red-team exercises showed indirect prompt injection paths through document retrieval and tool chains that could exfiltrate customer data. Regulators and the board expected accountable controls, not 'we trust the model.'
What Threesixty did
Ran agentic assurance assessment: injection surfaces, tool misuse, privilege escalation paths, not classic network pentest alone.
Deployed ClawGuard policies with gateway inspection and real-time violation feed to SecOps.
Established bi-weekly adversarial scenarios against production-like workflows with prioritised remediation.
Delivered leadership-ready reporting: blast radius, failed controls, and continuity options.
Technical approach
ClawGuard: policy management, per-agent assignments, agent-level tool validation, prompt sanitisation, gateway request/response inspection. Command Center violation feed with evidence export. Scenario-based red team aligned to AI Health Audit backbone; remediation tracked in audit log with Concordium-anchored evidence for high-severity events where configured.
Results
- Five hundred plus agents brought under centralised policy with violation evidence SecOps could action.
- All critical findings from initial assessment remediated within ninety days with retest sign-off.
- Board and regulator conversations supported by evidence packs, not dashboard screenshots.
- Reduced duplicate security tooling as regions consolidated on one governed stack.