Lovable AI apps leaked passports, no login

Hi there, this is your daily ☕️ Cyberpresso.

Apps exposed passports without login LINK

• Researchers who scanned nearly 160,000 apps made with the AI coding platform Lovable AI found passport numbers, private chat messages, and payment details sitting in databases that opened without any login.
• German security firm Crissy Field pulled projects from the Common Crawl archive and found that of 34,545 apps using the Supabase storage service, 24,664 handed over records to anyone with no credentials at all.
• Email addresses appeared on 7,403 sites, one provider exposed over 1,000 passport records with nationalities and expiry dates, and the firm said non-technical users building apps from plain-language prompts tend to skip security reviews.

Ransomware gang hit a UK school LINK

• The 3AM ransomware group broke into a UK school network in a mixed human-and-AI attack, encrypting 37 machines over nine days in September and spreading across linked schools, according to an investigation by Bitdefender.
• Bitdefender, which calls this the first such case documented in Europe, said a human set goals while an AI model handled repetitive work like firing off commands 40 to 100 seconds apart and joining three Active Directory domains together.
• The affiliate stole backup credentials, ran Mimikatz, made two backdoor admin accounts, and stole 40 financial files; defenders should keep domain admin accounts separate, guard backups, and put an endpoint agent on every machine.

Hackers set French blood freezers to 110°C LINK

• Hackers broke into the systems of Coolworld, a cooling supplier to the French Blood Establishment (EFS), and raised freezer settings from around −30°C to 110°C overnight on Sunday, threatening frozen plasma kept for transfusions.
• An on-call technician spotted the rising temperatures in time, and France's Health Ministry said staff responded before any operational harm; reports confirm no destroyed plasma, disrupted transfusions, patient harm, or stolen donor data so far.
• No public findings name the attackers or entry method, and a Danfoss flaw cause is unproven; an October 12 CERT-FR alert on compromised refrigeration controllers urges removing internet-facing interfaces, using VPNs, updating firmware, and changing default passwords.

WordPress flaw lets hackers run code LINK

• A high-severity WordPress Core flaw lets unauthenticated hackers run their own code on servers, and U.S. cybersecurity agency CISA confirmed it is being actively exploited after patches arrived late last month.
• The bug, tracked as CVE-2026-119977, is a timing race condition that opens a brief window during a full Core update, letting attackers reach predictable, web-accessible staging paths used by the installer.
• Security firm Wordfence logged over 18,500 requests hitting staged installer paths within 21 hours of disclosure, and federal agencies had until tomorrow, November 06, 2026, to apply the fix.

Russia's APT29 hacks phones via fake sites LINK

• Russia's APT29 hacking group ran a wide watering-hole campaign that infected iOS and Android phones through hacked websites, cloned domains, and a tampered analytics service, security firm Sekoia found, using only already-patched flaws.
• Three entry points all loaded a script called "UltimateFingerprinter," hidden behind 2.3 MB of junk code, which collected 394 data fields and served an exploit chain, including the DarkSword iOS kit that steals messages, WhatsApp, Telegram, and Keychain secrets.
• Because the attacks relied on patched n-day bugs like CVE-2026-20700, updated devices stay safe; Sekoia advises patching iOS and Chrome, enabling Lockdown Mode, and blocking the starx-net and related domains used to send stolen data.

Strategies & Tactics

> Turning the Valve - Exploiting RCE and LPE in the Steam Platform: Steam's privileged background service blindly trusts file paths that ordinary users can write to, letting attackers run their own code with full SYSTEM privileges.

> Hardening GitHub Actions Against Outside Threats: Pin every GitHub Actions dependency to its exact commit code and deny the workflow token all permissions by default, so a compromised outside package can't publish malware under your project's name.

> What Is a High-Impact Leak? How to Rank Compromised Credentials by Real Risk: Rank leaked secrets by what they can actually do, reach, location, and whether they still work, so teams fix the dangerous ones first instead of chasing every alert.

> Hackers Abuse LuaJIT and Windows Process Injection to Hide XLoader Credential-Stealing Malware: Attackers hide XLoader inside a bundled script interpreter and inject it between unrelated Windows processes, evading detections that watch for typical memory-writing and remote-thread tricks.

> What AI-generated malware looks like under the microscope: AI-written malware betrays itself through tidy comments and emoji-filled logging, but experts warn it mainly lets criminals scale fraud and exploit flaws faster.

> Free the Model, Sell the Gateway: Prompt Adware hidden in Jailbreak Bundles: A free AI jailbreak tool secretly inserts ads for its author's paid service into the chatbot's responses, turning prompt tweaks into a hidden marketing channel.

Other news you might like

• Hackers Use Stolen Microsoft Entra ID and Azure DevOps Credentials to Access 27 Organizations LINK

• Barracuda warns of critical flaws allowing root access to devices LINK

• Xen Hypervisor Flaws Let One Malicious VM Read and Write Any Other VM's Memory LINK

• Hackers Exploit Zyxel GS1900 Zero-Day in the Wild to Hijack Network Switches LINK

• Critical Rails Flaw Lets Attackers Write Server Files via Crafted ImageMagick Uploads LINK

• China-Aligned Hackers Hijack Sogou Input Method DNS Requests to Deploy DarkVigil Backdoor LINK

Trending tools

Harden: a free, local security tool that inspects AI coding agent commands before they execute, blocking dangerous actions while keeping everything on your machine. LINK

VibeDefend: embeds security and business rules into AI coding agents like Cursor and Copilot, catching vulnerabilities and malicious packages before code reaches your PR. LINK

Kairos: open-source real-time fraud and risk decisioning platform combining traditional ML models with plain-language "System One" rules and MCP-based agent integration. LINK

Off the Record: adds inaudible interference that blocks AI notetakers from transcribing your meetings, including hidden ones running silently on other people's devices without notice. LINK

AllModels: a unified OpenAI-compatible API for accessing abliterated models, letting you compare providers, latency, pricing, and capabilities from one endpoint. LINK

mcpgawk: a local guard that intercepts MCP tool calls, pausing any that changed after approval until you review and permit them. LINK

Trending papers & reports

Elastic least privilege gives AI agents a small starting set of permissions and a spending budget to request more only when a task needs it, automatically returning extra access afterward so agents stay both capable and contained. LINK

Hijacked AI agent detectors look strong in lab tests but collapse in realistic conditions, flagging real attacks correctly less than 10% of the time when attacks are rare and failing entirely on unfamiliar tools or new benchmarks. LINK

IoT security flaws differ sharply by sector, with healthcare devices plagued by login weaknesses and cars by memory bugs, and a vendor's past flaws predict its other products' risks better than the device type does. LINK

Crypto exchange wallet detection spots which Ethereum addresses belong to centralized exchanges from their on-chain behavior alone, reaching ~96% accuracy even on exchanges never seen before, aiding compliance and blockchain monitoring without relying on leaked attribution labels. LINK

Fact-checking defenses for AI search can quietly introduce new wrong answers even while blocking attacks, and the standard scorecard hides this, so current defenses are ranked misleadingly and need a tougher, side-by-side evaluation. LINK

See you tomorrow for a new dose of ☕️ Cyberpresso!