Securing AI Agents: Identity, Authorisation, and the DPACT Framework

The podcast discusses the evolving role of AI agents in software systems and the critical challenges surrounding their deployment, particularly in the areas of security, identity, and governance. As AI transitions from simple chatbots to autonomous agents capable of performing tasks with real-world impact, concerns arise about access control, authorization, and accountability. These agents often handle sensitive data and operate without direct human oversight, necessitating new frameworks for delegation that clearly define what actions they can take, under what conditions, and on whose behalf - without impersonating users.

A key focus is the development of structured governance models like the D-PACT framework, which emphasizes delegation, policy enforcement, auditability, context boundaries, and time-limited authority to ensure safe agent operations. Real-world examples illustrate risks such as overprivileged agents issuing unauthorized refunds or hiring systems being exploited through prompt injection, highlighting the need for guardrails and incremental implementation of controls. The discussion underscores a broader industry shift - from static access models to dynamic, context-aware authorization - and stresses that legacy systems designed for human users must evolve to support secure, auditable, and bounded agent interactions.

• What if you designed agent delegation with strict, time-bound grants instead of permanent keys?

  • Move: Replace any long-lived API keys used by AI agents in your apps with short-lived, task-specific grants (e.g., valid only for 15 minutes and scoped to one action like "read user preferences"). Use JWTs or OAuth2 with fine-grained scopes to enforce this.
  • Why Now?: The industry is shifting from static credentials to dynamic delegation models; early adoption reduces future rework and prevents security drift as agents gain more access.
  • Expected Upside: Lower risk of credential leakage, easier audit trails, and compliance readiness - especially under GDPR/SOC 2 - while positioning your product for agentic scalability.

• What if you implemented a human-in-the-loop checkpoint for high-impact agent actions in your SaaS product?

  • Move: Identify critical operations (e.g., refunds, data exports, purchases) and insert mandatory human approval steps via modal prompts or email confirmations before execution. Log all decisions with trace IDs.
  • Why Now?: As AI agents become 24/7 actors, unchecked automation can lead to costly failures (e.g., mass refunds); now is the time to build guardrails before scaling.
  • Expected Upside: Prevents runaway agent behavior, builds customer trust through control, and creates upsell opportunities for tiered autonomy plans (e.g., "Pro" users unlock automated approvals up to $50).

• What if you audited every AI agent in your system for identity separation and access scope today?

  • Move: Conduct an inventory: list each agent, its tools, data access, authentication method, and delegation model. Enforce identity separation by ensuring no agent uses shared human tokens or broad service accounts.
  • Why Now?: Most breaches in agentic systems stem from poor identity management - not model flaws - and governance maturity starts with visibility.
  • Expected Upside: Immediate reduction in overprivileged access risks, clearer compliance posture, and a foundation for policy-based frameworks like D-PACT that differentiate your product's safety in the market.

• Implement the D-PACT framework (Delegation, Policy, Auditability, Context, Time) to define clear boundaries and guardrails for any AI agent used in your software workflows.
• Conduct an inventory assessment of all AI agents in use, documenting their tools, data access, and credentials to identify and mitigate overprivileged or risky configurations.
• Use short-lived, context-specific delegated grants instead of long-lived API keys or human tokens when granting agents access to sensitive systems or data.
• Introduce audit logging with agent IDs, task IDs, and trace IDs to ensure accountability and enable post-action review of autonomous agent behavior.
• Apply autonomy tiers to agent actions (e.g., propose-only, read-only, reversible) and require human-in-the-loop approval for high-impact operations like financial transactions or customer data changes.