Video Summary - ECDSA's Curves Origins - When AI Becomes the Attacker

Summary

Microsoft reportedly fixed 1,411 vulnerabilities in October, following another unusually large batch the previous month. The hosts attributed the surge partly to Microsoft using its M-Dash AI system to find and remediate flaws in legacy code.
The reported categories included 677 privilege-escalation flaws, 317 remote-code-execution flaws, 254 information-disclosure flaws, 73 denial-of-service flaws, 40 security-feature bypasses, and 28 spoofing flaws. The update also addressed vulnerabilities described as actively exploited or publicly disclosed.
Microsoft introduced 15 experimental “risk cohorts” to group related vulnerabilities and help organizations assess and coordinate patching. The hosts noted that these groupings may help teams divide patch-testing work, but do not indicate on their own which flaws are most dangerous.
Unlike the prior month’s update cycle, which reportedly caused significant breakage, October’s patches had fewer observed consequences and fixed some earlier problems. One reported issue left the Smart App Control setting grayed out on some Windows 11 devices; a registry-based workaround was available.

Palo Alto Networks’ Unit 42 described an attack in which a human attacker used frontier AI agents to conduct a coordinated intrusion. The agents reportedly compressed work estimated to take multiple human red teams about two weeks into fewer than 10 hours, using more than 50 MITRE ATT&CK techniques.
After gaining access through a public-facing service, the agents mapped the network, searched code repositories for secrets, obtained privileged credentials, exploited CI/CD workflows, and stole cloud access keys. They also used the victim’s own cloud AI infrastructure to help continue the attack.
The attackers established overlapping persistence across areas such as SSH keys, cloud identities, serverless functions, containers, and CI/CD pipelines. They left behind an 80-page technical audit detailing weaknesses they had found and exploited.
Unit 42 recommended rapid, coordinated containment; inventorying and restricting AI endpoints and credentials; monitoring for unusual automated behavior; and strengthening DevOps protections, including mandatory multi-person code reviews and immutable branch protections.
Steve Gibson’s broader assessment was that AI-driven attacks can exploit organizational oversights—such as old credentials, archived messages, and forgotten infrastructure—not just software bugs. He argued that organizations need defenses that monitor both systems and user activity.

The show discussed a supply-chain attack involving Ledger hardware wallets sold through a reseller in Southeast Asia. A confirmed compromised device reportedly contained a hidden cellular modem and eSIM that read the wallet’s 24-word recovery phrase from the device display during setup and sent it to the attackers.
The device’s secure chip and software were not reportedly compromised; the attack targeted devices after they entered the distribution channel. The attackers then drained wallets automatically across several blockchains. Bitquery’s analysis tracked about $12 million from 335 wallets, while other researchers’ reported address totals suggested losses exceeding $86 million. The figures covered different sets of wallets and chains.
Ledger said it was investigating, asked affected sellers to stop sales, and reminded resellers that returned devices must not be resold. The program also mentioned reports of hardware implants in Trezor devices, while noting that Trezor disputed claims that its devices had been compromised.
Practical guidance discussed: Buy from sources you trust, be wary of devices that appear opened or include preprinted recovery words, and do not reuse a recovery phrase from a potentially compromised device.

Steve outlined two alternatives for people who want more control over a wallet’s hardware and firmware:

SeedSigner: Uses a Raspberry Pi Zero, camera, screen, and open-source firmware to handle QR-based transactions without a radio. The hardware and firmware can be assembled or built by the user. Steve cautioned that counterfeit Raspberry Pi boards exist and recommended established electronics suppliers rather than unknown marketplace sellers.
Blockstream Jade: A commercial device based on an ESP32 chip. The user can install the firmware, providing more hands-on oversight than simply trusting a sealed, preconfigured wallet.

Anthropic’s Frontier Red Team reported that an unreleased Claude research model traced the origins of the seed values used for standardized ECDSA elliptic curves in less than half an hour.
The model found that the seeds were generated using the 1990s-era lrand48 pseudo-random number generator, initialized from Unix timestamps, rather than being derived from the rumored hidden phrase or a hash.
It also identified methods used to generate some curve base points. The discovery was a reverse-engineering result, not a cryptographic break: Anthropic said it did not reveal a weakness or require users to change how they use ECDSA. Gibson said the explanation makes deliberate backdoors in the parameters less plausible.
Gibson emphasized the broader implication: an AI agent could solve a long-standing technical mystery from sparse clues, illustrating the potential power of similar systems when used by attackers.

Listener Tom Apalenic said AI coding tools let him build and test hobby projects much faster, and asked for more explanations of how AI harnesses, models, and tools work together.
He also suggested a continuous SpinRite-style utility that would refresh stale SSD sectors in the background. Gibson said he had similar ideas and was considering work related to SpinRite 6.1.
The episode also noted Anthropic’s offer to scan qualifying, opted-in open-source projects for vulnerabilities.

Steve Gibson — security analysis and commentary.
Micah Sergeant — guest host, filling in for Leo Laporte.
Reports and sources discussed included Palo Alto Networks Unit 42, BleepingComputer, Bitquery, Ledger, Anthropic’s Frontier Red Team, Crypto Briefing, and listener Tom Apalenic.