AI Nightmare - Hacking at 0 Hour

The speaker, a bug bounty hunter, starts from a view common in his community: AI can only find low-severity issues. Early in 2026 he tested that belief by building a Claude Code setup around his manual methodology, adding the program scope, test accounts and written instructions the model could follow. During a two-week live hacking event in Belgium, the setup produced a set of validated, unduplicated reports and a top placement, with no manual hunting on his side.

He then expanded the system with MCP servers, extra skills, sub-agents for triage and memory management. Examples included a sandbox escape in gVisor that gave access to the host, an account takeover chain across tenants built on an invitation flow, and an authorization flaw in a GraphQL gateway tied to agent creation. He also described failure modes: the model drifting out of scope, treating intended behavior as a bug, and overrating impact without business context.

His practical advice covered running tools in a sandbox, treating recon as context, mixing models, keeping notes, maintaining backups and controlling token cost. He argued that the time from patch to exploit is shrinking, so the human work moves toward managing context and verifying output. He suggested that organizations run AI code review and external attack surface monitoring, and that defenders need the same tools as attackers.