CISA just dropped an advisory about AI-generated malware

The advisory (ICS-ALERT-24-XXX) isn't theoretical. Researchers demonstrated a proof-of-concept where an LLM was prompted to generate STEP 7 / TIA Portal compatible code that manipulates memory blocks on S7-1500s, bypassing the standard "know-how protection" by exploiting how the PLC handles encrypted block downloads. The attack chain doesn't need a zero-day — it chains together legitimate engineering functions: upload station → decrypt block (using extracted keys from engineering workstation) → inject malicious OB/FC → re-encrypt → download. The AI component? It writes the malicious OB1/FC logic and the decryption routine in seconds, tailored to the specific firmware version pulled from the device's SSL certificate.

What makes this nasty: the generated code passes static analysis because it uses standard Siemens instructions (SCL/STL), no exotic opcodes. Only behavioral monitoring catches it — like unexpected DB writes to safety-critical memory areas (DB100+ typically reserved for safety programs).

Mitigations CISA lists, ranked by effectiveness:

• Network segmentation: isolate engineering workstations from PLC subnet (VLAN + firewall rules blocking port 102/ISO-on-TCP except from approved MACs)
• Firmware signing enforcement: enable "Secure Boot" on S7-1500 (requires FW ≥ V2.9) so modified blocks won't load
• Know-how protection + strong passwords: 16+ char, not the default "siemens" — the PoC brute-forced weak passwords offline
• Disable unused services: turn off web server, OPC UA, SNMP if not needed — each expands attack surface
• Integrity monitoring: hash critical DBs/OBs nightly, alert on change (simple Python script polling via python-snap7)

# quick integrity check example
from snap7.client import Client
from snap7.types import Areas
import hashlib

plc = Client()
plc.connect('192.168.1.10', 0, 1)
db_data = plc.read_area(Areas.DB, 100, 0, 256)
print(hashlib.sha256(db_data).hexdigest())
plc.disconnect()

The uncomfortable part: most plants still run S7-300/400 with no secure boot path. Upgrade cycles are 10-15 years. AI-generated exploits for those legacy platforms are trivial — no encryption to bypass, just plaintext blocks over MPI/DP.

Vendors (Siemens, Rockwell, Schneider) are adding AI-assisted anomaly detection to their SCADA suites now. But the cat-and-mouse game shifted: attackers iterate payloads in minutes, defenders wait for quarterly signature updates.

If you run critical infra, the advisory isn't optional reading. Patch the engineering stations first — that's where the keys live.