Model Guardrails Won’t Stop AI-Powered Attacks

The synchronized rollout of Anthropic’s Fable 5, Mythos 5, and OpenAI’s GPT-5.6 Sol marks a shift: AI companies now lead with safety classifiers, not just capability. This creates the impression that model-level guardrails are the primary defense.

Realistically, guardrails apply only to the good actors. Open-weight models like GLM 5.2 already match Mythos on vulnerability discovery, and criminal tools like WormGPT have always ignored safety. The result is an asymmetric field where defenders may be slowed by overzealous classifiers while attackers use ungoverned models.

The first claimed agentic ransomware, JADEPUFFER, was likely primitive but signals the inevitable. It exploited a default credential flaw in Langflow, moved at speed, and destroyed data. Whether agentic or not, the automation trend matters more than the label.

Engineering leaders should treat AI safety as an architectural problem. Locking down extension policies, hardening PowerShell execution, and monitoring runtime behavior are more durable than relying on vendor promises. The real lesson: every AI capability you adopt will also be adopted by adversaries, so invest in system resilience, not just model safety.