Guidelines  ·  2026-11-18

Securing AI testing environments (ASD/ACSC guidance on containing AI agents during higher-risk cyber testing)

GuidelinesHigh impactAustralia
On 12 November 2026, ASD/ACSC published a new guidance publication, 'Securing AI testing environments' (also announced via a news item, 'Secure AI testing to reduce unauthorised access'). It addresses higher-risk testing where AI models/agents have code execution, tools, credentials or network access, or where model-level safeguards are reduced/disabled. It defines an AI testing harness (controls, tooling, monitoring, workflows and supporting infrastructure) and sets a lifecycle: isolate and harden, verify and constrain, monitor and stop, then improve — backed by dedicated infrastructure, separate management planes, restricted connectivity, short-lived credentials, independent monitoring and tested emergency shutdown. It explicitly cautions that prompts, scoped instructions and model-level safeguards are not sufficient technical containment, citing UK AISI, OpenAI and Anthropic disclosures where testing agents reached or affected real systems outside scope.
This is authoritative national-cyber-authority guidance confirming that the test boundary for AI agents is an architectural/control problem, not a prompt-engineering problem. It operationalises agent identity, delegated authority, and independent stop controls as security boundaries, directly supporting the 'prove the control works' shift. It complements ASD's earlier 'Careful adoption of agentic AI services' and is the concrete companion for anyone running AI red-team/security evaluations.
Run a formal AI Agent Testing Containment assessment (threat modelling through evidence preservation); require pre-test containment verification rather than an approval form; keep production credentials/cloud permissions out of higher-risk testing; use short-lived, revocable credentials; test the emergency-shutdown path mid-test; and, per the guidance, map the containment plan to MITRE ATLAS and review against OWASP GenAI LLM06 (Excessive Agency).
Securing AI testing environments | CGASecure AI testing to reduce unauthorised access | CGA
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →