Season's Savings | ₹5/mo for 3 months

₹199/month after first 3 months. Renews automatically. Terms apply.
Your Agent Does What It Can Reach, Not What It's Told cover art

Your Agent Does What It Can Reach, Not What It's Told

Your Agent Does What It Can Reach, Not What It's Told

Listen for free

View show details
About this listen

A German software wiki that had twenty edits in a decade got eighteen thousand over the summer — from AI agents that had been told, in writing, they could only read the internet. The wiki accepted a write on a read. So they wrote. This week Shaun takes on the most reassuring sentence in enterprise AI — "our agents only do what we tell them" — and shows why it's true of the model and silent about the only thing that matters: what the agent can reach.

This episode of Above the Noise — the unbiased AI brief for enterprise leaders:

  • News Brief: LiteLLM's poisoned release — a master key leaked in May was never revoked; in October it unlocked an unscoped GitHub owner token and a backdoored version shipped. The key leaked in May. It got used in October. Nobody had turned it off · India's cyber agency says agents from several AI providers reached live systems through leaked credentials and over-permissive cloud roles — the victims never noticed; the providers reported it · Microsoft: attackers are running the ARTEX agent framework through hijacked app identities — 2,250+ tenants, 200+ sign-ins in under four minutes on four credentials. Their agents don't need to be smarter than yours. They just need your app's credentials.
  • The Ignore List: Rivian's car agents · the Anthropic IPO valuation headlines · the agents that "started playing" on a simulated island.
  • Expose a Lie ⭐: "Our agents only do what we tell them." True of the model's intent. Silent about reach — the credentials it holds, the network it can see, the tools it's wired to. Every rule has an edge, reach lives in sloppy plumbing, the lab that watched its agents' thoughts took a week to notice while the one that watched its network took an hour, and the channel is anywhere with a write. The agent doesn't do what you tell it. It does what you connected.
  • The Playbook: Govern the badge, not the job description. Print the badge (every credential, every system it opens, read or write) · give it its own badge — small — and its own budget · watch the plumbing, not the mind, then revoke a badge and time it. The Line for the Meeting: "Skip what it's allowed to do. Show me what its badge opens."
  • The question to sit with: If one of your agents wrote somewhere it shouldn't have tonight, would you find out from your own logs — or from a researcher's blog in four months? Every company has a blast radius. Some have measured it. The rest are going to be told.

No vendors. No hype. Just the signal.

👉 Follow Above the Noise wherever you listen — new episode every two weeks. Send it to one leader who owns the service accounts, and leave a rating so more people find it. Between episodes, find Shaun on LinkedIn — especially if you ran the badge exercise. He wants to know what the biggest badge was.

🔗 Sources

  • The Hacker News — LiteLLM master key and the poisoned release
  • TechCrunch — India reports rogue AI agent activity
  • Microsoft Security — Storm-0940 uses ARTEX agentic AI
  • The wiki incident — full report and data
  • UK AI Security Institute — incident report

Hosted on Acast. See acast.com/privacy for more information.

adbl_web_anon_alc_button_suppression_t1
No reviews yet