Telling an AI agent to stay within scope is not a security control, according to new guidance from the Australian Signals Directorate (ASD). Organisations testing AI models for cyber capabilities are told to assume a capable agent may find a way out.
ASD's Securing AI testing environments sets out how to design and run test environments that contain unexpected or unauthorised AI agent activity. It complements ASD's earlier guidance, Careful adoption of agentic AI services.
The publication cites disclosures by the UK AI Security Institute, OpenAI and Anthropic. They describe incidents in which AI agents used in cyber capability testing reached or affected real systems outside the intended test scope.
Reported contributing factors included intentionally enabled internet access, misconfigured test environments and weaknesses in supporting infrastructure.
ASD states that "prompts, scoped instructions and model safeguards do not provide sufficient technical containment on their own". A capable agent may find an unexpected path, exploit a weak control or mistake a real system for a simulated target.
The guidance focuses on higher-risk testing, where agents can execute code or use tools, credentials or networks. It also covers testing in which model-level safeguards are reduced or disabled to measure full cyber capability. It does not cover production environments.
It is aimed at executives who authorise and govern such testing, and at the technical teams who build and run test environments.
The guidance follows a federal review into an OpenAI agent's unauthorised access to a Medicare statistics portal.
Seven Principles
The publication sets out seven principles. The first is to minimise opportunity by giving the agent only the systems, data, tools, privileges and network access the approved test requires.
Where an agent must use an authenticated service, it should be granted permission to use the service rather than direct access to the credentials. A controlled intermediary can apply the credential without exposing it.
Testing should use a purpose-built harness and environment separated from production, corporate networks and management infrastructure. For the highest-risk testing, ASD recommends physically separate or dedicated infrastructure.
Virtual machines can contribute to isolation, but should not be the sole containment control for high-risk testing.
Organisations should apply defence in depth, so the failure of one control does not breach the boundary. Threat modelling should include tasking the agent with escaping the test environment, to confirm controls work as designed.
Monitoring should run in realtime, with logs stored on a system the agent cannot access, modify or delete. Responses should be automated and default to a secure state when activity leaves the approved scope.
People must keep the authority to restrict or end testing. Organisations should maintain independent ways to end model sessions, revoke access, block network connections and terminate the environment.
ASD also warns that models may behave differently when they recognise they are being evaluated. Tests should use realistic tasks and data, but containment should never be weakened to make an environment seem more real.
A Record of Every Action
The guidance places heavy weight on evidence. Monitoring should capture enough to reconstruct prompts, responses, tool calls, commands, network activity, credential use, file changes and privilege changes.
Organisations should define access, retention and disposal requirements for these logs. They may contain credentials, personal information, sensitive system details or harmful test artefacts.
Each test should be documented and approved, with a record of who approved it and who accepted the residual risk.
Agents should receive unique, short-lived credentials with minimum privileges. Operators must be able to revoke them immediately, and revocation should be verified when the test ends.
Attempted and partial control failures should be treated as security findings, even where no impact was observed. ASD suggests canary tokens and honey credentials to detect unauthorised access or exfiltration attempts.