AI
The agentic attack surface is being built faster than the defences around it
Three disclosures this week describe one shift. Censys found over 800 Artex deployments reachable without a password; Apiiro counted 2,539 live secrets in 1,122 public repositories exposed by saved agent approvals; OpenAI's own pipeline surfaced 948 Linux privilege-escalation flaws for $1.2 million. OpenAI then gave Hacktron expanded model access to probe its third-party integrations. The thread is custody, not capability: agents hold credentials, touch production systems and leave configuration behind. The organisations that cope are those that can inventory what their agents can reach.