Threat Model: Agent-to-Agent Delegation Abuse
When AI agents delegate tasks to each other, the delegation chain becomes an attack surface. How to threat-model and contain A2A delegation abuse.
AI agent security, identity, governance, cost, and the engineering behind the control plane.
When AI agents delegate tasks to each other, the delegation chain becomes an attack surface. How to threat-model and contain A2A delegation abuse.
HMAC signatures plus timestamp replay windows are the minimum bar for secure webhooks — here's why unsigned endpoints are dangerous and how to fix them.
You cannot govern agents you do not know about. OpenTelemetry ingest turns existing instrumentation into an agent inventory. Learn what the ingest path has to solve, the tenant-isolation trap, and how discovery becomes governance.
Agent spend is forecastable if you model it as volume times cost per task rather than extrapolating a total. Learn the decomposition, the non-linearities that break naive projections, and what to plan for.
Version AI agent workflows like code, diff changes between snapshots, and roll back safely when a new version causes regressions or runaway costs in production.
Agent loops resend the same context on every step, which is exactly what prompt caching is for. Learn how to structure prompts for cache hits, what invalidates them, and the security considerations.
A dead-letter queue nobody watches is a place where work goes to be forgotten. Learn what belongs in a DLQ, how to alert on it, and how to replay safely once the cause is fixed.
A practical incident response runbook for AI agent breaches: contain damage, revoke credentials, investigate with tamper-evident audit trails, and recover.
Over-broad MCP tool permissions give attackers an amplified attack surface. Learn the failure modes and control classes that shrink the blast radius.
Any administrative action a single person can both request and approve is not controlled. Learn where maker-checker belongs in an AI platform, how self-approval sneaks back in, and how to keep it usable.
Public agent benchmarks are useful for narrowing a shortlist and misleading as a selection criterion. Learn what they measure, how contamination distorts them, and how to benchmark against your own workload.
Detect and redact PII before it reaches AI models or persists in logs — covering entry points, detection techniques, redaction strategies, and compliance.
Agents that transact need a spend mechanism with reservations, atomicity, and idempotency. Learn the ledger design that makes agent payments auditable and the concurrency bugs that make them wrong.
At-least-once delivery plus agents that take real actions equals duplicate side effects. Learn where idempotency keys belong, why retries are a cost multiplier, and how to make replay safe.
How signed trust manifests and scoped admission controls let organizations share AI agents across boundaries without exposing data or credentials.
A staged rollout playbook for AI agents: inventory risk, run a scoped pilot with guardrails in place, define go/no-go criteria, and expand on evidence.
How agent trust scoring models aggregate identity, behavior, and attestation signals into a runtime gate that controls what autonomous agents can do.
Contain what AI agent tools can do: tool-level scoping, allow-lists, dry-runs, and human approval gates for high-consequence irreversible actions.
Design abuse-resistant rate limits for AI agents: choose the right unit, window shape, and enforcement scope to protect costs and downstream systems.
Choose the right enforcement action for AI agent guardrails — block, redact, or warn — and understand the fail-open vs fail-closed security trade-off.
What makes an audit trail credible to an auditor or court: hash-chaining, per-row digital signatures, and external anchoring explained for engineering teams.
Keep PII out of AI agent prompts, responses, and logs using detection-and-redaction controls that satisfy GDPR, HIPAA, and audit requirements.
Keep API keys and credentials out of agent prompts and source code. The four pillars of secrets management: storage, delivery, access control, and rotation.
RBAC governs who can configure agents; ABAC governs what agents can do per request. Learn which model fits each authorization decision on an AI platform.