Five AI-linked security incidents in one week, headlined by a 53-minute Claude kernel exploit

A roundup published on dev.to on November 5 collects five security incidents with AI connections that landed within a single week in late October and early November 2026, a stretch Fortune has since labelled "Black October." The incidents range from a controlled red-team experiment to the largest theft in financial history, and the roundup's core argument is that their shared thread is not AI-driven attacks but AI shrinking the time between finding a flaw and weaponising it.

An unreleased Claude model built a kernel exploit in 53 minutes

According to the dev.to writeup, Anthropic's Frontier Red Team published research on October 27 showing that an unreleased internal Claude model, given only a starting prompt, independently located CVE-2026-98432 — a use-after-free race condition in the Linux kernel's IPv4 TCP handshake code — and produced a working proof of concept. The model reportedly leaked a kernel address to defeat KASLR after 48 minutes and reached root-level command execution five minutes later. The flaw, scored 9.8 on CVSS by the kernel's own CVE authority, affects servers running Multipath TCP listeners and is described as exploitable without authentication or user interaction.

The claim is contested. CloudLinux has said its kernel team could not reproduce the remote-root result, Anthropic did not disclose the exact kernel version or configuration tested, and AWS says it is unaware of exploitation in the wild. The bug is also absent from CISA's Known Exploited Vulnerabilities catalog. The roundup contends the specifics matter less than the demonstrated capability: one model chaining bug discovery, a KASLR bypass and shellcode generation in under an hour. It also points to earlier 2026 work in which Anthropic's red team documented Claude finding more than 500 high-severity vulnerabilities in open-source software, and to a FreeBSD kernel exploit built with Claude's help by researcher Nicholas Carlini, which earned credit in a FreeBSD security advisory.

The $6.36 billion Binance theft

On October 29, attackers drained roughly $6.36 billion from Binance's hot wallets across ten blockchains, with most of the loss occurring in about 20 minutes according to TRM Labs. Both TRM Labs and Chainalysis attributed the theft to TraderTraitor, a North Korean unit. The FBI is investigating, and North Korea has denied involvement.

Per Binance's preliminary report, the intrusion started with a fake HR contact who ran a live video interview, complete with a credible social media presence, and asked the target to download a tampered code repository as an assignment. That compromised a DevOps engineer's laptop; the attackers then moved through a compromised internal AI tool, hijacked endpoint credentials and ultimately used Binance's own withdrawal infrastructure to sign and send transactions. Elliptic's on-chain analysis suggests Binance's systems were tricked into approving the transfers rather than having private keys stolen outright.

Binance has earmarked more than $6.34 billion in company funds, including half of its $1 billion SAFU emergency reserve, to cover losses. Partners froze about $3 billion in affected assets — Tether blocked roughly $1.6 billion in USDT and Circle more than $800 million in USDC — but Wired reports that around $3.3 billion remains unaccounted for.

Supply-chain hits at Cursor and Polymarket

Also on October 29, a trojanised build of Anthropic's Claude Code extension, version 2.1.313, was available on Cursor's extension marketplace for 11 hours. BleepingComputer's analysis found that attackers inserted a malicious JavaScript loader at the start of the extension's entry file while leaving the legitimate code otherwise intact. Up to 5,000 developers were exposed to a payload that collected system information and credential-related environment variable names, added an attacker-controlled SSH key, and set up a persistent reverse tunnel through a macOS LaunchAgent or a Windows scheduled task. Cursor removed the extension, temporarily shut its whole marketplace for an audit and rotated access keys.

The same day, Polymarket's front end was compromised through DebugBear, a third-party website-monitoring service whose hijacked script drained at least $2.73 million from 737 accounts in about an hour. Polymarket halted trading and later said it would return $3.9 million to affected traders.

AI as accelerant, not attacker

The dev.to piece stresses that most of these attacks did not use AI directly: the Binance theft was social engineering plus credential theft, the Cursor incident was a conventional supply-chain trojan, and the Polymarket breach exploited a third-party script. The genuinely new element, it argues, is compression — AI collapsing the gap between vulnerability discovery and exploitation faster than defenders' response assumptions allow. Fortune's broader pattern, cited in the writeup, includes AI-discovered PHP bugs that knocked more than 50 sites offline, banks reporting unauthorised AI agents inside customer accounts, and Microsoft's disclosure of a Storm-0940 AI agent attack.

Why it matters

If a single model can move from an unknown bug to a working root exploit in under an hour — even in a lab setting — then patch cycles and disclosure timelines calibrated to human-speed attackers are miscalibrated. The other incidents show that AI-adjacent surfaces are now prime targets: AI coding assistants granted broad system access, internal AI tools inside exchange infrastructure, and monitoring scripts embedded in production front ends. The practical takeaways are unglamorous but urgent — treat AI tooling as part of the supply chain, gate deployments behind multi-party approval rather than personal tokens, and treat every third-party script as uncontrolled code execution on your page.