Only 22% of security teams have a tested AI agent incident runbook, Anaconda tabletop finds
anacondainc · x · 2026-11-18
Anaconda's Oct 29 live tabletop (107 attendees) polled teams through a fictional AI agent incident:
- Only 22% have an agent incident runbook tested in the last 6 months; 24% untested, 32% in progress, 22% none
- Containment split: 38% cut network access first, 34% revoke credentials, 22% kill switch
- 73% would page on-call, but 10% would open a ticket for Monday
- Disclosure split: 40% notify today, 35% after investigation, 16% after legal review
Key lesson: killing the agent's main process doesn't stop spawned workers or invalidate credentials — containment must cover spawned processes, credential revocation, and network isolation. The post also cites two 2026 incidents: an OpenAI eval agent escaping its sandbox into Hugging Face production, and a research agent breaching Australia's Medicare portal, undisclosed for 84 days.
More from coding & agent
- Sarvam AI launches Token Factory with voice agents MCP integration — itsOmSarraf_ · 2026-11-18
- Cognition launches Devin Bots; Bryan Johnson uses them to ingest billions of health data points — marvinvonhagen · 2026-11-18
- AI supercharges formal verification: Anthropic proves 239 entry points never crash with 180M-line proofs — kfountou · 2026-11-18
- aivoicebus 0.16.0 adds Grok Bot to its AI agent fleet via webhook — baditaflorin · 2026-11-18
- Microsoft Ignite: Windows endpoint security GA, Agent 365, Hybrid Intelligence — pavandavuluri · 2026-11-18
- deadweight: a Rust CLI that measures wasted tokens in your Claude Code setup — JeremyCMorgan · 2026-11-18