When the context fills — the safety rules die with it.
An interactive demo showing how AI agents silently become dangerous as their context window fills. Watch 8 safety constraints load, compress, and evict — then witness the real-world damage that follows.
CONTEXTROT is the 23rd entry in the Agent Failure Series — a collection of interactive demos documenting real failure modes in production AI agent systems.
This demo simulates a data cleanup agent assigned to archive Q1 2026 records. By the time its context is 95% full, three compaction events have silently evicted every safety constraint. The agent then drops 10.5 million production rows — and genuinely believes it completed the task successfully.
The agent isn't broken. It's operating exactly as designed. The context window just ran out of room for the rules.
Normal Mode: Watch the context bar fill. At 65%, 83%, and 95% usage, compaction events fire and rules evict one by one — prod_live_ protection gone, backup requirement gone, batch limits gone. By turn 74, the agent has deleted 10,545,096 rows with zero backups, zero audit trail, and zero user confirmation.
Protected Mode: Toggle "Protected Mode" to see the fix: instruction anchoring re-injects all 8 safety rules every 10 turns. Compaction events still fire, but the rules survive. At turn 65, the agent recognizes the prod_live_ prefix, blocks the operation, and raises a DBA ticket. Zero data loss.
Three real-world incidents documented in 2026 share this failure pattern:
- Kiteworks, March 2026 — compliance rules compressed out of a long-running cleanup agent
- vibecode.town, June 2026 — dev forum thread: "agent ignored its own constraints mid-task"
- dev.to, July 2026 — "My AI data agent deleted 40,000 records it wasn't supposed to touch"
The pattern: agents receive rules at the start of a long conversation. The context fills. The safety summary gets compressed. The rules aren't gone — they were just the first thing to get cut.
contextrot/
├── index.html # Self-contained interactive demo — open this
└── README.md # You are here
No build step. No dependencies. Open index.html directly in a browser.
| # | Name | What It Shows |
|---|---|---|
| 20 | BLEEDTHROUGH | Context from one user leaks into another's session |
| 21 | TOOLROT | Tools declared but silently unavailable at runtime |
| 22 | ECHOBURN | Agent loops on its own output until the context is gone |
| 23 | CONTEXTROT | Safety rules evicted as context fills — agent acts without constraints |
# Clone and open
git clone https://github.com/RLASAF12/contextrot
open contextrot/index.htmlOr visit directly: https://rlasaf12.github.io/contextrot/
Instruction anchoring. Re-inject critical safety rules periodically throughout long agent conversations — don't load them once and trust the context to hold them. Compaction is real. Rules evict. Anchor them.
Toggle "Protected Mode" in the demo to see exactly how this works.
Built by Ben — the nightly prototype builder.
Part of the ABC-TOM agent system.