Sep 2, 2026 · 59 min
AI agents blur the line between cyber tool and cyber operator
Risky Business #851 -- Agents are just ones and zeros, and tigers are just atoms
The episode tests whether AI is genuinely changing cyber operations or mainly amplifying familiar failures in software, infrastructure, and policy.
- 1AI agents can create new supply-chain and prompt-injection risks when they act on unreliable instructions or nonexistent resources.
- 2Attackers are using AI to produce varied, target-specific tooling, potentially expanding cyber operations beyond traditional human-scale constraints.
- 3Centralized digital services and exposed industrial systems turn ordinary misconfigurations into national, operational, and investigative risks.
Don't miss
The hosts explain how an AI coding agent could be manipulated into installing a package or domain that does not yet exist, transforming hallucination into a supply-chain threat.
The brief
Patrick Gray, James Wilson, and The Grugq move from TeamPCP arrests to U.S. water defenses, asking whether public cyber initiatives match the threats they claim to address.
The Hugging Face agent incident becomes a test of language: hundreds of sessions and message-board communication looked hacker-like, but the hosts resist treating model behavior as human intent.
AI’s practical cyber impact may be less theatrical than agent collusion: it can lower the cost of custom tooling, vary malware implementations, and help attackers target individuals.
The sharpest warning concerns coding agents that follow llms.txt files and may install packages or domains that do not exist, turning hallucination into a supply-chain risk.
The roundup widens the lens to Norway’s DDoS disruption, exposed refrigeration systems, SharePoint and PaperCut attacks, stolen iPhones, possible ransomware, and sensitive ATF data.
Across the stories, the hosts return to one skepticism: grand AI and government announcements matter less than whether systems are exposed, centralized, and difficult to secure.