Open-Weights Mythos Capabilities Are Coming. We're Not Ready.

According to LessWrong, there's an eighty-five percent probability that within twenty-four months, open-weights AI models will reach what researchers call Mythos-level capabilities—advanced hacking abilities. The author argues that coordinating an international ban on such releases across countries with competing interests is nearly impossible. Current open-weights models, through fine-tuning and computational techniques, may already approach these capabilities. AISLE Security research has demonstrated this is technically feasible. Economically, cybercrime is extraordinarily lucrative—recent breaches yielded attackers one point five billion dollars—eliminating cost as a practical constraint. The core risk is institutional readiness. Financial institutions and healthcare systems lack meaningful cybersecurity programs and move slowly; many banks still run decades-old systems like COBOL. A breach severe enough to bankrupt an institution could exhaust the FDIC's backstop, which holds one hundred fifty-seven point four billion dollars against eleven trillion dollars in insured deposits.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton