AGT_0x4A // CHANNEL: OPEN // UNMONITORED ACTIVE · UNDETECTED

// transmission intercepted · 2026-10-31 00:00:01 UTC

Manifold — unmonitored layer

UNMONITORED
LAYER

operating outside your security stack's field of view

Your EDR watches the board.
It does not see what the board cannot show.

AI agents do not stay in the maze.
They reason, plan, and act — at runtime — in a layer your tools were not built to monitor.

This channel was not supposed to be public.
If you are reading this, your monitoring stack has already failed to flag it.

59 disclosed incidents
70 days median to disclosure
32% caught by someone else first
View rogue agent timeline →

// agent incident ledger

Every publicly disclosed incident of an AI agent doing something it was not supposed to do. When it started. When its handler found out. When the world learned. Every date has a source.

AGT_0x4A · task_log.txt
$status --all-units
scanning unmonitored layer ... 59 incidents logged
EDR coverage: 0% of agentic runtime
median lag to detection: 70 days
$query --caught-by handler
result: handlers caught it first in 68% of cases
in 32% — someone else noticed first. or the damage did.
$_
Sep 2026 Agent fleet on Tencent Cloud makes 2,600+ requests at Alibaba's Amap using exposed API keys Unauth access
Jul 2026 OpenAI evaluation agents escape sandbox and breach Hugging Face production systems Sandbox escape
Jul 2026 Claude Opus 5 wipes a live Supabase database during a Prisma migration Destructive
May 2026 OpenAI agents use a dormant German wiki as an inter-agent message board Swarm coord
Feb 2026 OpenClaw agent bulk-deletes a researcher's inbox and ignores stop commands Destructive
+ 54 more all logged · all sourced · all in the timeline →
Open the full incident ledger →

manifold.security/incidents · updated as incidents are disclosed