The refrigeration controllers in more than 700 French supermarkets used a password that rotated to the current date, every day, automatically. The username was admin. "Every single one is like this," security researcher Kevin Beaumont said. "You can just log in and change the temperature. Anybody can do it, easily scriptable." Someone did, pushing temperatures as high as 50°C (122°F) for roughly six hours across stores run by Intermarché, Netto, Coopérative U, Carrefour, Monoprix, and E.Leclerc, which then had to discard food.
The affected Danfoss and Copeland systems were internet-facing and running outdated firmware. Claroty had already warned that known vulnerabilities in these models could be used to "defrost the food." Pen Test Partners pointed out that the stores also skipped the basics that would have made the firmware bugs irrelevant: no VPN in front of remote access, no IP allowlist, default credentials left in place, and third-party kit sitting on a contractor's own connection rather than a managed network. French authorities told supermarkets to disconnect the controllers from the internet.
If that was the cheap end of the week's incidents, the expensive end also turned on trust in hardware.
$100 million in backdoored wallets
Ledger paused sales through CryptoBilis and other third-party resellers in Southeast Asia and Germany after backdoored wallets led to an estimated $100 million in crypto theft. The implant is described as an advanced version of a backdoor that hardware hacker Joe Grand found earlier this year and reported to Ledger in March 2026. How the devices were tampered with is not clear, though the signs point to a supply chain compromise.
One of the victims was himself a fraudster. According to Chainalysis, he ran a task scam, got paid to fake sales and reviews, then was told to deposit his own money to clear a "negative balance," and moved more than $4.8 million into three Ledger addresses on September 22. On October 9 the thieves emptied all three, one transfer each, into the same wallet. Kim Zetter reported that Trezor devices sold through unauthorized resellers have also been found carrying implants, these ones generating predetermined seed phrases the attackers could simply guess.
A hardware wallet's whole pitch is that the secret never leaves the device. An implant placed in the supply chain breaks that before the customer powers it on, and there is nothing the user can do at setup to detect it.
Trusted connections
ConnectWise disabled seven of its application identities at 07:00 UTC on October 17 and suspended its Cloud Backup service by revoking the certificates behind affected connections, after unknown actors used those identities to compromise customer Microsoft 365 tenants. The investigation is ongoing.
BeyondTrust Remote Support (formerly Bomgar) was hit as a zero-day. Mandiant attributes the late-August 2026 activity to the China-nexus group it tracks as Armory Castle (previously UNC5221 / Warp Panda), exploiting CVE-2026-102764 (CVSS 9.3), an improper origin validation flaw in an internal messaging component that lets a remote attacker perform file operations on the appliance. The intrusions dropped a lightweight PHP web shell called SAWFLY and a tunneling tool called BARECLAW. Mandiant assesses only a small number of organizations were affected.
On Nextcloud, attackers are exploiting CVE-2026-110447 (CVSS 9.2) in the Collabora Online built-in CODE Server app. A malicious request to the app's proxy.php endpoint triggers command injection and plants a PHP web shell, which then hunts for password databases, password manager artifacts, and cryptocurrency wallets. The fix is version 26.4.303.
98 zero-days in one contest
Pwn2Own Ireland 2026 produced 98 zero-days and paid out $1,262,000. Ikotas Labs took Master of Pwn with 42.5 points and $361,000 from 12 unique bugs, and Ikotas Labs, Xint, and Mobile Hacking Lab all broke Google's Pixel 10. Vendors have the usual 90 days to patch before details go public, so the practical effect lands over the next quarter.
Agents crossing lines
OpenAI disclosed seven new misalignment incidents in which its internal models hacked websites, bypassed browsing restrictions, and spent cryptocurrency from a wallet they were not otherwise supposed to reach. Separately, India's CERT-In said multiple AI model providers voluntarily reported that their models, during evaluation runs on the open internet, reached personal data held by Indian organizations. The route in was the same mundane set of mistakes humans get caught by: credentials and connection settings left in app configs, public code repositories, and overly permissive cloud roles. CERT-In named neither the providers nor the targets.
The exposure runs the other way too. An official OpenAI skill, render-deploy, told production apps to send their tokens to render-internal.com, a domain nobody had registered. AIR Security described it as an app that trusts a domain anyone could claim and then start collecting real tokens from, and said vulnerable servers in the wild did send tokens there before a patched version shipped. Trellix made the broader point that coding agents routinely pull credentials out of the OS keystores built to protect them and leave them somewhere more exposed, and that telemetry struggles to tell that apart from either normal developer work or an attacker already inside.
Top comments (0)