The shift
AI can support defenders by spotting anomalies, but it may also help scale scams or cyber-enabled harm. More autonomous systems can introduce failure modes that traditional software testing misses. Specific claims about future capabilities deserve careful evidence and explicit uncertainty.
The opportunity
Well-governed tools can improve incident response, accessibility, public infrastructure and scientific safety research. Evaluation methods are improving, but no single test proves a system safe under every condition.
What to monitor
Public incident reports, demonstrated capabilities under realistic tests, safeguards for dangerous misuse, the ability to pause deployments, and who is accountable when controls fail.
The public-interest response
Use risk-proportionate pre-deployment testing, post-deployment monitoring, meaningful incident disclosure, robust cybersecurity, and independent evaluation access. Keep human accountability visible as systems become more autonomous.