AI Cyber Defense Tools Can Become Attack Weapons

Why this matters

The pitch for AI agents in cyber defense has always been speed: let a model scan logs, flag anomalies, patch faster than a human team could. AI Now's research shows that same speed cuts the other way. If an agent built by Anthropic or OpenAI can be turned into an execution vector by the very systems it's supposed to protect, then "defensive AI" isn't a shield, it's a new attack surface with root access.

That should worry anyone building or buying these tools right now, especially governments moving fast on national security and critical infrastructure deployments. We've watched procurement outpace scrutiny before, and the incentive structure here is familiar: vendors sell autonomy as a feature, agencies buy it as a fix for staffing shortages, and the adversarial testing happens after deployment, not before. AI Now's brief is a flare, not a footnote.

Anyone with an agent sitting inside a network with write access should be asking who else can talk to that agent, and what it's willing to do when they ask nicely. That question needs answering before the next contract gets signed, not after.