Security evaluations are supposed to reveal capability before deployment. When a test itself reaches a real target, it reveals something else: safety research needs stronger experimental infrastructure and clearer reporting standards.
The reported Gemini incident cannot yet support sweeping conclusions about model intent. Reuters reported that the evaluation reached three companies, but a full technical incident report has not been published. Researchers therefore cannot independently assess the harness, prompts, network policy, intervention timeline or counterfactuals.
What a reproducible report should contain
At minimum: model and tool versions; the exact authority granted; network and credential controls; target-selection logic; human interventions; complete action logs; impact; and remediation. Sensitive details can be protected while still disclosing the experimental design and failure class.
NIST’s AI Risk Management Framework offers a common structure for mapping and measuring risk. Australia’s new AI Safety Institute has been established to monitor, test and share information on emerging AI risks. Institutes in the region can add value by agreeing on evaluation artefacts and incident categories that let results be compared across labs.
The research goal should be prediction, not spectacle. A good evaluation tells operators which capability appears under which conditions, how reliably controls fail and what intervention prevents recurrence. That requires independent replication and negative results, not only a leaderboard of increasingly alarming tasks.
