This episode breaks down:
Plus Bishop Fox research on how agent harness design shapes the results of AI security tools, and why AI-generated vulnerability findings are cheap to produce but costly to verify.
The $6.36 billion Binance hack began with a fake job interview, according to Binance, which says North Korea-linked attackers spent months compromising employees and internal systems before draining exchange wallets across 10 blockchains on October 29, 2026. Blockchain intelligence firm TRM Labs attributes the attack with high confidence to TraderTraitor, North Korea's elite hacking unit. It is the largest crypto theft on record, about four times the $1.5 billion Bybit hack of February 2025.
How did attackers get inside Binance? Binance's chief security officer said a software engineer was contacted in September through a legitimate job platform by someone posing as a project director. After a video interview with a fake HR representative and a second interview with a purported CEO, the engineer was told to download a coding assignment onto his work computer. It contained zero-day malware that evaded endpoint security. In a separate incident in June, another engineer installed a tampered Homebrew package while searching for an AI tool. Binance replaced that device, but says the wallet team remained exposed. On October 20, attackers used the machine compromised through the interview, along with an internal AI tool, to reach two wallet engineers' computers. On October 29, they used stolen credentials to deploy a container and run a malicious script, and funds began leaving at 20:05 UTC.
What is EIP-7702, and how was it used in the Binance hack? EIP-7702, introduced in Ethereum's Pectra upgrade in May 2025, lets an ordinary wallet hand control of its behavior to a smart contract by signing an authorization, and the delegation stays in place until the wallet signs a new one. The attacker submitted one transaction containing 50 authorizations signed by a Binance hot wallet, each pointing it to a sweeper contract that only the attacker controlled. Each authorization is valid only at a specific transaction count, or nonce, and a busy exchange wallet's count keeps moving. Signing 50 consecutive counts in advance guaranteed that one would match when the transaction executed, and one did. Three minutes and 24 seconds later, the wallet sent 167,464 ETH and 63 tokens, including 1.615 billion USDT, to the attacker, who kept sweeping new deposits for hours afterward. Academic researchers had described this delegation-based phishing vector in a December 2025 paper.
Why did wallets on other blockchains send funds to the attacker too? In the same window, Binance wallets on the XRP Ledger, Bitcoin, Tron, Solana, and five other networks sent funds to the attacker through ordinary signed transfers, with no delegation involved. These wallets use different chains, key types, and signing schemes. In TRM's assessment, that points to the attacker gaining the ability to make Binance's hot wallet signing system sign transactions of its choosing. Because Binance's hot wallets use multiparty computation, no single key could revoke the malicious delegation, and containment took until 00:10 UTC on October 30.
Who is behind the Binance hack? TRM attributes it to TraderTraitor, the same group the FBI named for the Bybit hack. Binance said the FBI has attributed the attack to North Korea. The campaign matches earlier North Korea-linked heists: small Tornado Cash withdrawals funding an operator wallet, staging that began well before the theft, and signatures obtained in advance, then executed in a single burst.
How much of the stolen crypto has been frozen or recovered? Binance says about $2.85 billion has been frozen and about 98% of the affected tokens have been replenished with company funds, so customers won't absorb the loss. TRM counted roughly $2.7 billion frozen onchain by stablecoin issuers Tether, Circle, Ripple, and World Liberty Financial, plus Arbitrum's Security Council. The attacker still controls more than 420,000 ETH and 18,270 BTC, about $2.4 billion in assets beyond any issuer's reach.
What indicators should crypto exchanges and treasury teams watch for after the Binance hack? Review signing and endpoint telemetry for:
Mitigation: Configure signing policies to reject EIP-7702 authorizations unless they are expected, and keep a tested, pre-approved method for signing a revocation. Separate signing infrastructure and approval paths by chain and wallet tier, so one compromise can't produce valid signatures across 10 networks. Treat the corporate devices of privileged engineers as part of the external attack surface, and test internal AI tools for what they can reach.
Sources: TRM Labs, CoinDesk, arXiv
CVE-2026-41709 is an insufficient-logging flaw in Broadcom VMware ESX that lets a user with administrator access perform certain operations without those actions being recorded in the audit log. CrowdStrike's 2026 Legal Threat Landscape Report says the China-nexus espionage group Warp Panda exploited it as a zero-day in intrusions at several U.S. organizations, including law firms, as late as early 2026. Broadcom rated the flaw Low, with a CVSS 3.1 score of 2.7, and fixed it in advisory VMSA-2026-0006 on July 29, 2026. CISA added it to the Known Exploited Vulnerabilities catalog on October 28, with a November 11 deadline for federal civilian agencies.
The low score reflects the high privileges the flaw requires. For an attacker who already holds hypervisor credentials, erasing the forensic record is the payoff, and that makes incident response on an affected host extremely difficult to scope. Because exploitation began before the patch, organizations that patched in July may still have unlogged activity in the months before it. Patch ESX regardless of CVSS, review whether low- and medium-severity SLAs leave hypervisors waiting, and consider a retrospective hunt across the pre-patch window.
Source: SecurityWeek
CVE-2026-111585 is a high-severity integer overflow (CVSS 8.8) in the remote-control screen processing of TeamViewer clients on Windows, macOS, Linux, and Android. Discovered by security firm Remedio, it lets a malicious or compromised host crash the viewer of the support technician who connects to it, and under some conditions execute code on the technician's machine. TeamViewer has released fixed versions, and no exploitation in the wild has been reported.
CVE-2026-97631 is a critical, 16-year-old buffer overflow (CVSS 9.5) in the PXE Base Code driver of TianoCore EDK II, the open-source firmware foundation used by UEFI vendors including AMI, Insyde, and Phoenix, as well as Microsoft's Project Mu. While a server boots over the network, the driver copies ICMP error messages into a fixed 504-byte field without enforcing that limit. An attacker with a foothold on the local network can use this to run code before the operating system loads and before Secure Boot or measured boot has verified anything. Ask your firmware supplier about fixes and disable network boot where possible.
Both flaws target the tools IT teams rely on. A technician's workstation carries credentials and management access to every system it supports, so run remote support from dedicated jump boxes and treat unsolicited sessions as untrusted. For PXE, compare your imaging network's actual reach against its as-built design, since a single port left spanned after a one-off build is enough access for this attack.
Sources: SecurityWeek, Security Affairs
An unreleased Anthropic AI model found security flaws in Chrome, Safari, Android, Windows, and Linux and chained them into five working proof-of-concept exploits, according to a research post from Anthropic's Frontier Red Team. The Chrome chain combined four separate flaws, from V8 memory corruption to a use-after-free in Chrome's FedCM sign-in feature, to run code outside the browser sandbox on a stock Mac within seconds of a page loading. The Linux chain ran commands as root over the network against servers running a multipath TCP listener. Every flaw was reported and patched before publication, including fixes in Chrome 156 and iOS 26.5.2.
The work was done in a lab after a researcher prompted the model, and Anthropic does not break down how much each chain depended on human involvement or name the model. Even so, the range of targets is the signal: chaining low-severity bugs into a working path has historically been specialist work, and skilled exploit developers using these tools will move much faster. Defenders should assume the time between a cluster of moderate findings and a working exploit chain is shrinking, and prioritize remediation accordingly.
Source: The Hacker News
Subscribe to our PODCAST
Real talk on the threats, trends, and tactics shaping security today
Recommened Resources
Download
Your download is starting in a new tab. If it does not start automatically, use the button below.
