Why AI Does Not Need to be Innovative to be Dangerous

Summary: Bitdefender's Martin Zugec argues that LLMs are optimized to predict the most statistically probable output, making them poor at finding the low-probability anomalies hackers exploit, and that probabilistic hallucination breaks malware that requires determinism - a ransomware encryptor that fails even 5% of the time destroys the payment-for-restoration trust the RaaS economy runs on. He contends AI's real advantage is industrialization: driving the marginal cost of an attack down so a $5 attack against a $500 ransom pays off, with standardized environments such as default Microsoft 365, standard AWS configurations, and uncustomized SaaS platforms as the near-term hunting ground.
What's compelling: The limiting factor for cybercrime has been human bandwidth, and AI removes it - which means the small organization that assumed it was too small to target now sits in a commodity infrastructure pool that automated agents can sweep profitably at $500 a victim.

Read the original at The Hacker News ↗