Sony breach exposed 2.6M customer records

Hi there, this is your daily ☕️ Cyberpresso.

Sony reveals breach exposing 2.6 million records LINK

• Sony Network Communications says attackers broke into the So-net member database and stole personal data on more than 2.58 million current and former customers, plus the email passwords of 54,401 users.
• The stolen records cover So-net members and plala subscribers, exposing names, phone numbers, email addresses, birth dates and connection IDs, along with 16 bank account records, though no passwords from membership data or credit card details leaked.
• The breach hit on October 19 and was confirmed on October 23; So-net cut off the attacker's server, reset the exposed email passwords yesterday, and warns customers to watch for scam messages that exploit the leaked details.

Hackers automate attacks with 44 AI agents LINK

• Hackers turned CyberStrikeAI, an open-source penetration-testing platform, against mostly Japanese organizations, running 44 coordinated AI agents that handled network scanning, exploitation, persistence, and data theft without a person steering each individual task.
• Over a short stretch in October 2026 the tool checked weaknesses at 19 organizations, 16 of them in Japan, assigning work to specialized agents, logging findings in a database, and even producing Japanese-language fake login pages.
• TrendAI found the activity in exposed attacker files; no new flaw was used, entry began through an unauthenticated admin interface on a legacy system, and many reports recorded no breach, with deep compromise hitting only a few targets.

Oracle business software flaw is under active attack LINK

• Hackers have started probing a critical Oracle E-Business Suite flaw that lets them run commands without logging in, with threat intelligence firm Defused spotting exploratory attempts against its decoy systems yesterday.
• Tracked as CVE-2026-101767 with a severity score of 9.8, the bug lets a crafted web request trigger SQL injection through the Thin Client Framework servlet, chaining into command execution under the EBS application account.
• Oracle patched the hole on Tuesday last week before the probes appeared, so these are attempts, not confirmed breaches; admins should patch now and meanwhile pull the servlet from the URL firewall allowlist.

Malware hid in oversized GitHub files LINK

• Researchers found a wave of malware hiding inside oversized GitHub files, part of the testmetest npm campaign that turned 42 booby-trapped repositories across 11 real owners into tools for stealing credentials and publishing bad packages.
• The payloads measured about 13 MB each, big enough to slip past GitHub's search and content-retrieval checks, and used a loader mixing character scrambling, AES-128-GCM decryption, and a Bun bootstrap to launch the credential thief.
• SafeDep reported the worm now runs when an app simply imports the package, so turning off install scripts won't stop it; separate workflows attempted CVE-2026-41940 exploitation against cPanel, though JFrog confirmed no successful downstream compromise.

UEFI flaw lets attackers run code LINK

• A 16-year-old flaw in TianoCore EDK II, the open-source UEFI firmware that boots most of the world's servers, lets attackers run their own code before the operating system and its security checks load.
• Tracked as CVE-2026-97631 with a severity score of 9.5, the buffer overflow sits in a PXE network-boot driver that copies ICMP error messages past a fixed size limit, letting an attacker on the local network corrupt memory and run code at the CPU's highest privilege level.
• There's no sign it has been exploited or used for initial outside access, but AMI's early fix last month exposed details before January's planned disclosure; firms should patch firmware, turn off network boot where possible, and watch logs, since Secure Boot and TPM don't block it.

Strategies & Tactics

> Scalable Vector Grief: Six of eight widely used SVG cleaners failed to block hidden scripts, proving that accepting image uploads can still expose sites to code-injection attacks.

> How LLM-Based Penetration Testing Harnesses Work: A harness is the code that executes an AI model's text instructions as real actions, turning a passive text generator into a tool that can actually probe websites for security flaws.

> SBOMs Alone Can't Reveal the Scope of a Supply Chain Attack: A software inventory can be outdated and miss hidden dependencies, so base incident response on what actually runs in production, not static lists.

> This tool can recover deleted AI chats from your phone – how it works: A research tool called ORISA recovers fragments of deleted local AI chats from phone memory, helping cybercrime investigators but potentially exposing private conversations if misused.

Other news you might like

• Hackers Compromise Camera Management Software to Deploy Malware on Windows Computers LINK

• Veeam Patches Critical Backup & Replication Authorization Bypass Vulnerability LINK

• Fake ChatGPT, Claude apps spread malware via signed AI binaries LINK

• GoAhead TLS Flaws Allow Revoked Client Certificates and Enable Session Decryption LINK

• Anthropic's Unreleased AI Model Chains Flaws Into Working Chrome, Safari, Windows and Linux Exploits LINK

• Adobe Commerce PolyShell Attacks Breach Hundreds of Stores as Brotli Exploit Traffic Emerges LINK

• FortiBleed 2.0 Reveals Thousands of Fortinet VPN and Active Directory Credentials in Exposed Configurations LINK

Trending tools

Google Gemini 3.8 Flash and Cyber: fast, affordable models for agentic workflows, multi-step reasoning, coding, and cybersecurity tasks like vulnerability detection and automated patching. LINK

Ghostifier: scans your inbox headers to find who holds your data, then unsubscribes, opts you out of sales, and sends tracked deletion requests automatically. LINK

Replay QA Security Scan: runs automated pentests against your web app to catch injection flaws, broken access control, IDOR, and cross-tenant data exposure. LINK

face2social: reverse image search that matches a photo across Instagram, Facebook, LinkedIn, TikTok, and X to verify identities and catch scammers. LINK

chat-recall: unifies AI assistant histories from Claude Code, Codex, Cursor, and OpenCode into one searchable log, strips secrets locally, and flags leaked keys. LINK

Koreshield: screens customer messages, retrieved documents, and proposed tool calls in AI support agents, catching prompt injection and unsafe actions before execution with full logging. LINK

Trending papers & reports

Invisible text watermarks that tag AI writing by swapping look-alike characters work far better for some languages than others, needing ~2 characters to detect English but ~93 for Lao, leaving low-resource languages harder to trace. LINK

Model theft detection gives proprietary AI owners a reliable way to prove a rival copied their model through its public interface alone, catching imitators that current watermark checks miss when outputs are generated cautiously. LINK

Prompt injection defense teaches an AI assistant to read untrusted content without obeying hidden commands buried inside it, cutting one automated attack's success rate to 10% versus 97% on an undefended model. LINK

AI agent safety gates decide per action whether to run, verify, or ask a human, cutting harmful side effects in simulation from ~15 percent to ~7 percent without spamming humans with approval requests. LINK

Cross-bank fraud detection shows banks can spot money laundering that spans multiple institutions using transaction records they already hold, with sharing just two extra counts per account adding value while keeping customer data private. LINK

See you tomorrow for a new dose of ☕️ Cyberpresso!