Somewhere in your company, someone tired of a ticket backlog has probably already given an AI agent a login. Silas says give it the keys. Daniel says you’ve built a privileged account that reads text written by strangers and takes actions. Vivian brings the documented cases, the counter-evidence, and the one test that tells you which of them is right about your setup.
The office scene in this episode is a hypothetical composite, built from the mechanism of two documented incidents. It is not a real company.
In this episode
- The two documented agentic intrusions — one state-sponsored campaign in which the AI did 80–90% of the work, and one cloud break-in where the agent needed no exploit to escalate
- How attackers used AI coding assistants already installed on developers’ machines to harvest 2,349 credentials from 1,079 systems
- The honest counterweight: Mandiant says 2025 was not the year breaches directly resulted from AI, and only 1.3% of AI-found vulnerabilities have been confirmed exploited
- Why the question isn’t “does the agent have admin rights” but “what is the worst thing this identity can do, and how fast can you turn it off?”
- The five-minute revocation test, the “lethal trifecta”, and a three-sentence policy that fits on a sticky note
- A hypothetical week for an office IT manager: the fast setup, and the slower one that doesn’t end in an incident report
About the show
All Eyes takes one question the industry is arguing about, puts two people who genuinely disagree in a room, and does not let either off easy. Hosted by Vivian Calloway, with Daniel Ramos and Silas Wray, who argues the other side for a living. New episodes most Tuesdays.
🔊 This podcast uses AI-generated voices. Every voice you hear is AI-generated, including the host. Daniel Ramos's voice is his own, cloned with his written consent. We disclose this at the start and end of every episode. Read exactly how we use AI →
Sources cited on air
- Anthropic — Disrupting the first reported AI-orchestrated cyber espionage campaign (GTG-1002, November 2025)
- Sysdig Threat Research — agentic threat actor and container escape (May 2026)
- The Hacker News; GitGuardian — the Nx “s1ngularity” supply-chain attack (August 2025)
- Aim Security / CVE-2025-32711 — EchoLeak, zero-click prompt injection in Microsoft 365 Copilot (June 2025)
- Mandiant (Google Cloud), M-Trends 2026 — AI and 2025 intrusions
- VulnCheck — exploitation of AI-discovered vulnerabilities; AppOmni and Cloud Security Alliance — the Salesloft Drift OAuth breach
- Daniel Ramos, The 2026 InfoSec Playbook, Chapter 7 — Using and Securing AI (Intelligent Automation, 2026)