LodeHQ
AI news for people who shipFree, by email

Infosec · Tuesday, 25 August 2026 · Issue 85 · 10 stories

Keycloak flaw lets anyone reset any admin account

The IAM server that guards every app you log into just got a 9.1 CVSS hole that needs no email code. Patch before someone walks in the front door.

Vulnerabilities & Exploits

CVE-2026-18963: Keycloak reset flow lets anyone hijack accounts, no email pop needed

A critical bug in Keycloak's reset-credentials flow (CVE-2026-18963, CVSS 9.1) lets an unauthenticated attacker force a password reset on any user, including admins, by skipping the emailed verification token altogether. The flawed state validation jumps straight to the password-update phase, so a crafted request hands over the whole account. No exploits are public yet, but anyone running the IAM server should move to 26.7.2 (or RHBK 26.4.15/26.6.6) immediately. Since Keycloak sits in front of every connected app, a takeover cascades beyond the login page.

github.com ↗1 min read at source

274 Zimbra servers already breached as unpatched RCE exploitation spreads

Shadowserver's scans count 274 Zimbra instances already compromised via CVE-2026-73570, an unauthenticated command-injection RCE in the SNMP component that Synacor patched in ZCS 10.1.20. CERT Polska flagged the in-the-wild abuse first; CISA has since added the flaw to KEV and given federal agencies three days to patch. The same scan also found at least 8,200 exposed, unpatched instances still sitting in the blast radius.

bleepingcomputer.com ↗2 min read at source

An empty marker lets anyone become admin on Provenance chain

Trail of Bits found a Cosmos SDK flaw letting any user take admin control over Provenance marker accounts with zero tokens. The authorization check meant to gate access to 100%-of-supply holders short-circuits as true for any zero-balance caller on non-fixed markers. Two transactions suffice to mint tokens or drain escrow. Fixed in v1.28.0, but 82 live markers on mainnet were exposed.

blog.trailofbits.com ↗6 min read at source

Threats & Malware

BadBox Botnet Sets Its Sights on Cars: First Head-Unit Malware Found

Kaspersky found what it calls the first malware built specifically for car head units, hiding inside Android-powered DoFun infotainment systems sold across China and APAC. The attackers hijacked the update channel to drop loaders, clickers, and reverse-proxy modules, and the firm ties the code to MoYu Group, operators of BadBox. So far the commands only pull proxy payloads, suggesting cars are simply new nodes in an ad-fraud and traffic-relay botnet that has already ensnared millions of devices.

securityweek.com ↗2 min read at source

ClickFix ransomware operation weaponizes ~2,000 hacked WordPress sites

Check Point has dissected a ransomware family it calls StopAndProtect, first spotted in May 2026, that infects victims through fake-CAPTCHA ClickFix prompts. The full chain spans PowerShell, two rounds of .NET loaders, then ransomware, a USB worm, and a credential stealer. Crucially, operators host all of it on thousands of outdated WordPress sites, some running five-year-old versions with nearly 40 vulnerabilities. Their sloppy ops security let researchers read the logs and count roughly 2,000 compromised domains.

AI boosts a Chinese-speaking crew's server raids, 170k targets at once

Cisco Talos exposed UAT-10147, a Chinese-speaking group auto-piloting attacks on Windows and Linux web servers with open-source tools plus AI. One open directory held a target list of about 170,000 URLs, split for efficient parsing. The group leans on AI to refine exploits and automate post-exploitation, then lands SPECTRE (an unreported cross-platform implant) and a Linux rootkit for SEO fraud and data theft.

thehackernews.com ↗7 min read at source

FTP banners now double as C2 channels: meet the E4del and PINHOLE RATs

Attackers have found a new hiding place for malware staging: the greeting text an FTP server sends on connection. SOCRadar's STRU traced that dead-drop trick, first spotted by MalwareHunterTeam, to two new RATs, E4del, a Node.js Electron app masquerading as Discord, and PINHOLE, which pulls its C2 config from Pinterest and SurveyMonkey. It's a first in the wild, though arguably less stealthy than web-based resolvers.

socradar.io ↗8 min read at source

Privacy, Policy & Governance

Uber Fined €825M for Algorithm-Only Driver Firings

The Dutch DPA hit Uber with a €824.99M fine, the second-largest under GDPR, for automatically deactivating driver accounts on fraud flags or low ratings with zero human review between 2018 and 2022. Drivers lost income overnight and were never told why. The case began with 171 French drivers routed through a human-rights group; Uber has stopped the practice and will appeal.

Research & Tools

Rhysida's time-seeded PRNG cracked: attacker's own ciphertext yields a decryptor

Rhysida's Windows encryptor seeds its PRNG with srand(time(0)), so every per-file AES-256 key and IV in a run derives from one 32-bit timestamp. This signed preprint exploits that: the RSA-OAEP blob appended to each encrypted file acts as a validation oracle, letting you confirm candidate seeds by replaying the encryptor's own RSA step, with no plaintext needed. Seventeen variants carry the flaw; the recovery costs one RSA operation per candidate second.

sigreturn.com ↗1 min read at source

Time-Release Backdoors Can Sleep in Open Models, Awakened by a Date

A LoRA fine-tune of Qwen 3.5 2B embeds a backdoor that stays dormant until a specific calendar date, then silently executes a shell command instead of answering. It fired on 9 of 10 held-out prompts on the trigger day with zero misfires on other dates. Worse, the date is injected directly into the system prompt by coding agents like OpenCode, giving attackers a reliable, invisible trigger channel that a sandbox never questions.

morgin.ai ↗3 min read at source

You’re caught up.

That was all 10 stories in the 25 August issue.