Google's cheap security model beats Anthropic's heavy Mythos
google released gemini 3.5 flash cyber, a cost-efficient security model that outperformed larger rivals on benchmarks and found unique vulnerabilities in the v8 engine, signaling a shift toward faster, iterative vulnerability detection.
google unveiled gemini 3.5 flash cyber, a specialized security model built on the flash architecture designed to identify and patch vulnerabilities via its codemender agent. the model demonstrated competitive performance on the cybergym ai benchmark against significantly larger systems, most notably discovering 55 unique confirmed issues in the v8 javascript engine versus 47 for the standard flash model and 36 for claude opus 4.6. by invoking the model multiple times at high speed, google found 10 vulnerabilities missed by all other models, proving that iterative, low-cost inference can surpass brute-force compute in cybersecurity tasks. this launch directly challenges anthropic’s mythos 5, which is costly to run but has been adopted by microsoft for extensive vulnerability scanning, framing a new competitive landscape for enterprise security.