STATION 02 // THREATS

The machines became economic actors. So did the attacks.

Dossiers below are illustrative reconstructions of public incidents and the attack surface the autonomous economy now exposes. Atlas treats every incident as a multi-source intelligence problem, not a single signal.

$3.4B

Crypto theft, 2025 record

≈500%

AI-enabled scam growth YoY

≈80%

Orgs running autonomous AI without real-time visibility

INCIDENT DOSSIERS // ILLUSTRATIVE

INC-PI-0512·prompt-injection takeover

Prompt-injection of an AI agent via X

May 2026. About $150K to $200K siphoned from a Base wallet. No smart-contract bug. The agent was simply talked into it.

SOURCES // agents · social · onchain

CONFIDENCE // 0.92

ATTRIBUTION GRAPH

INC-LR-0419·router exfiltration

26 malicious LLM routers identified

Researchers found 26 malicious LLM routers in circulation. One alone drained about $500K from connected agent wallets.

SOURCES // agents · autonomous

CONFIDENCE // 0.95

ATTRIBUTION GRAPH

INC-OP-0316·operational social engineering

Drift exploit reframed as social engineering

The $285M Drift incident was operational social engineering, not a code bug. Exactly the surface single-source code scanners miss.

SOURCES // social · onchain

CONFIDENCE // 0.97

ATTRIBUTION GRAPH

PULL QUOTE
An agent without an intelligence layer watching it is an unattended wallet.