Home/ Incident database/morris-ii-ai-worm-2024-03
Data breach Data exfiltration verified

Morris II: researchers demonstrate a zero-click worm that propagates between GenAI agents via self-replicating prompts

Cornell Tech / Technion (research) · Mar 5, 2024 · General agent

What happened

Researchers published a paper (arXiv:2403.02817, submitted March 5, 2024) demonstrating an AI worm dubbed 'Morris II', after the 1988 Morris worm. Crafted as an adversarial self-replicating prompt, it exploits RAG-based GenAI applications — demonstrated against Gemini Pro, GPT-4 and LLaVA — to trigger a chain of indirect prompt injections that spread the payload to other agents with zero clicks, using email-assistant spam and data-exfiltration as the demonstrated use cases.

Loss / impact
Mar 5
2024
Data
Failure mode
General
Agent type

Causal vector

Adversarial self-replicating prompt injected via RAG — worm-like propagation between connected GenAI agents

Source

Reported by arXiv (Cornell Tech / Technion). Verified against the primary report.

Read the original report ↗

How a spend firewall would have helped

A worm that spreads by making agents exfiltrate data is stopped by the same control as every exfil chain: outbound network calls from the agent must pass an allowlist enforced outside the model. Worm-infected or not, an agent whose egress is policy-gated cannot beacon an attacker's server.

The six rule types that contain this class of failure

Per-transaction cap

Any single spend above your ceiling is BLOCKED before it moves.

Daily total

Cumulative spend across all agent calls, bounded per day.

Velocity limit

Stops runaway retry loops — the #1 cause of overnight losses.

Merchant allowlist

Only approved destinations can ever receive funds.

Category rules

Flag high-risk classes (crypto, infra, refunds) for review.

Approval threshold

Above a value, the action waits for a human.

Related incidents

Don't be the next entry

Every incident in this database is the result of trusting a prompt, a provider cap, or a human review cycle. sipi.bot replaces all three with one deterministic call. 67 documented failures, one control.