Google's Wiz and the AI firm Calif used models from OpenAI and Google to find the flaws in August, according to reporting. OpenAI has said its model found one of them and has released a checking tool. Separately, Microsoft disclosed that a China-aligned group has used an AI tool since September against hundreds of organizations. The name "Black October" comes from security commentators and social media, not an official designation. The same window brought other incidents. LiteLLM, a widely used tool that routes requests to AI models, was briefly poisoned with password-stealing code on October 21 after attackers stole engineers' authentication tokens. Railway reported suspected access to customer secrets. A GitLab flaw was rated 10 out of 10. Browsers revoked trust in some certificates, and about 30,000 need replacing. Figures vary across reports, and several investigations are still open.
The episode matters because it shows AI compressing the gap between disclosure and exploitation. Major AI labs warned about these capabilities earlier in 2026, and Anthropic's limited rollout of its Mythos model in April was an early signal. Experts expect more fundamental weaknesses to surface in the coming months. Attorneys should confirm whether clients run PHP and whether patches are applied, check whether they used LiteLLM or Railway, and track certificate replacement deadlines. Any breach analysis should account for the short window between the first probing on October 17 and disclosure.