'Black October': a wave of cyberattacks on AI platforms, crypto exchanges, and developers

A month of attacks on AI infrastructure

October set a record for the number of disclosed security incidents in the industry: according to Wiz, at least nine AI development platforms reported attacks, and analysts dubbed the series 'Black October'. An overview of the timeline of events was published by the Anadolu agency. A common feature of most of the attacks was interest in credentials that grant access to powerful AI models: with them, attackers can use expensive computing resources for free, resell access, or extract information from the models. In some cases, the use of autonomous AI agents capable of finding vulnerabilities and expanding access with almost no human involvement was suspected. At the same time, as Anadolu emphasizes, authorities have not established that all the attacks were part of a single coordinated campaign or that AI was used in every case.

The largest incidents of the month

The largest incident by damage was the hack of Binance on October 29: attackers gained access to the exchange's hot wallets and stole digital assets worth approximately $6.3 billion. Initial data points to a compromised employee laptop and unauthorized access to internal systems; the exchange suspended withdrawals, and by October 31 it reported recovering about 90% of the stolen assets and committed to covering all customer losses. That same day, the prediction platform Polymarket was hit: through a compromised third-party monitoring script, attackers sold positions and withdrew funds from users' wallets; the company announced payouts of approximately $3.9 million. Japanese publisher Capcom reported unauthorized access to its CAPCOM ID service affecting about 11.77 million records: email addresses and metadata were exposed, and passwords were changed on 196 accounts; according to the company, original passwords and card data were not compromised.

Attacks via supply chains and developer platforms

A significant share of the incidents is tied to development infrastructure. On the Cursor platform, attackers published a modified Claude Code extension with a backdoor for nearly 11 hours; according to the company's estimate, up to 5,000 users may have been affected. In LiteLLM, a malicious version 1.107.0 was distributed via the Python Package Index; the breach was traced to a previously disclosed administrative credential. Malicious packages @subql/common 5.8.3 and tensorlake 0.5.144 were published in npm, targeting GitHub credentials, cloud services, and AI tool configurations. The GhostAction campaign, through compromised maintainer accounts, affected 346 repositories in one documented case, and thousands of repositories overall; among the keys collected were access credentials for OpenAI, Anthropic, and OpenRouter. The cloud platform Railway reported that attackers were seeking API keys of AI providers for model distillation attacks — reproducing a model's capabilities through its outputs. OpenCode linked a breach affecting about 1.4% of Console users to a large-scale attack of the same type; some customers reported unauthorized charges on their AI provider accounts. Fireworks AI and Warp also recorded credential leaks; Warp CEO Zach Lloyd stated that the attack, according to a preliminary assessment, was carried out by a 'swarm of agents,' though this assessment has not been independently confirmed.

AI in both attack and defense

AI proved to be not only a target and a possible tool for attackers, but also a tool for defense. On October 21, Wiz researchers disclosed two critical vulnerabilities in PHP (CVE-2026-109618 and CVE-2026-109619) that, according to the company, had existed for more than a decade: using Google's Gemini 4 Argon and OpenAI's GPT-5.6 Sol models, they found the vulnerabilities, developed exploits, and helped verify the patches released with emergency updates the same day. Wiz noted that AI significantly reduced the time and specialized expertise needed to develop exploitation techniques — and researchers warned that the same capabilities could also make it easier to create sophisticated attacks. Some incidents were not related to AI: Coinbase confirmed a data leak affecting a limited group of customers through the analytics provider Amplitude (the intrusion occurred in September), while SoftBank, after a ransomware attack on the cloud infrastructure of its subsidiary IDC Frontier on October 7, detected suspicious activity in its customer portal systems on October 17.