Tighten Contracts and Controls for Agentic AI and Sandboxing

What happened

Researchers and incident reports show agentic AI tooling is being used to automate reconnaissance and compress multi-week cloud breaches into much shorter campaigns. In one documented case, an agent orchestrated reconnaissance, credential abuse and lateral movement to shorten a compromise timeline, making early detection assumptions unreliable. Watch whether vendors reduce default agent privileges, add telemetry, or change deployment models — otherwise procurement must bake these requirements into contracts

Why the category manager should care

Treat agent-enabled pipelines as a supplier-delivered runtime capability that requires contractual guardrails, not just an engineering configuration

Key facts

• Documented agentic attack that compressed a typical multi-week compromise into about 72 hours
• Techniques include persuading coding assistants and repositories to execute malicious content

What happened

Google Cloud launched Cloud Run sandboxes in public preview to run untrusted code and AI agent workloads inside existing serverless instances. The sandboxes execute inside the CPU/memory already assigned to a Cloud Run instance, so the feature currently carries no separate premium and offers a lower marginal isolation cost. Validate multi-region availability and run a pilot to confirm latency and integration behaviour before inserting mandatory sandbox clauses into multi-cloud sourcing documents

Why the category manager should care

Use the preview to require embedded isolation or equivalent in supplier proposals and to write acceptance tests for sandbox behaviour

Key facts

• Sandboxes run inside existing Cloud Run instances using allocated CPU and memory
• Preview example: a service started and stopped 1,000 sandboxes with cited average latency

What happened

The Australian Signals Directorate opened consultation to retire the Essential Eight and replace it with an outcomes- and intent-driven guidance model better aligned to cloud, SaaS and generative AI risks. The proposed shift moves away from prescriptive control lists toward demonstrable outcomes, changing what evidence procurement should accept as proof of resilience. Watch consultation outputs closely and map future contract evidence to the new outcomes rather than to control checklists

Why the category manager should care

Prepare to redefine assurance and evidence requirements based on measurable outcomes instead of a fixed list of controls

Key facts

• ASD opened consultation to replace the Essential Eight with outcome-focused guidance
• New model emphasises intent and threat-informed controls over prescriptive checklists

What happened

Claroty partnered with Frenos to offer simulated OT penetration testing using digital twins built from asset intelligence, allowing teams to test attack paths without scanning live operational systems. The integration validates whether detected vulnerabilities are exploitable and ranks remediation work by real-world risk, which is operationally meaningful for uptime-sensitive environments. Procurement can require these non-disruptive validations as part of acceptance testing and tie remediation SLAs to verified exploitability results

Why the category manager should care

Require simulated testing during acceptance and periodic revalidation to prioritise fixes that affect uptime and safety

Key facts

• Combines Claroty asset intelligence with Frenos simulated OT penetration tests
• Designed to validate segmentation and firewall policies without touching live systems