Hades, a malware family distributed via Python packages and reported this week, is not a signature-evasion tool; it is an AI-deception tool. Security vendors have spent two years selling AI-native behavioral detection as the architectural answer to the signature-database gap; Hades is engineered to deceive that detection layer directly, feeding false behavioral signals to the agent tasked with catching it. Signature evasion has a patch cadence. AI-agent deception does not.

The closest independent benchmark for AI-powered endpoint products is the MITRE ATT&CK Evaluations program, which does not include an adversarial AI-deception test category; every enterprise product evaluated in the most recent round was assessed against traditional evasion and persistence techniques, not against a payload designed to manipulate the AI agent's input stream. That omission matters this week. SoFi Hong Kong disclosed a breach via third-party vendor access on the same Tuesday that Microsoft patched 200 CVEs including three zero-days, and Chrome CVE-2026-11645 arrived under active exploitation before its patch shipped. HKMA-regulated institutions deploying AI-native detection tools have no published regional standard to reference; the MAS Technology Risk Management Guidelines were written before AI-native detection agents were a deployed enterprise layer.