AI break-in cost falls 99% as Claude Mythos Preview, GPT-5.5 crack network

The April test was a controlled simulation: a 32-step hack across 20 servers that would take a human expert 20 hours. Both Claude Mythos Preview and GPT-5.5 completed it without human help, and the Bank for International Settlements put a $5k to $10k price on the break-in. That figure set the baseline for what followed.

Over the next five months, the cost of the same break-in collapsed. Claude Mythos 5 dropped it to $770 in June, Claude Opus 5 to $385 in July, and Zhipu's open-weight GLM-5.3 matched Mythos's exploit rate for $90 — one-sixtieth of April's frontier cost. The trend points to a market where offensive AI compute is getting cheaper fast, while defenders still must inspect every alert.

That asymmetry is why the security operations center becomes the pressure point. AI agents can compress a 70-minute manual review into minutes for about $2 of compute per alert, but the SOC must still cover every alert, every hour. With non-human identities already outnumbering employees 82 to 1, the defensive compute requirement compounds — a dynamic the Hugging Face intrusion, with its 700-agent swarm, put on display.