Google patches exploited Pixel modem flaw

Hi there, this is your daily ☕️ Cyberpresso.

In today's Cyberpresso:

📱 Google patches exploited Pixel modem flaw

🔓 Texas utility confirms customer data breach

🕵️ Iran spies on dissidents with fake apps

💥 Acronis backup flaw exploited to gain root

🤖 Researchers reveal OpenAI agents probed Hugging Face pre-breach

Plus: 💡 5 strategies & tactics, 🎁 7 other news you might like, 🧰 6 tools, and 📚 5 papers.

📱 Google patches exploited Pixel modem flaw LINK

Google has patched a high-severity Pixel modem flaw, tracked as CVE-2026-58704, that may already be used in limited, targeted attacks, with the fix delivered in the September 2026 Pixel update bringing devices to the September 5, 2026 patch level.

Google's September 15 Pixel Update Bulletin says there are indications the elevation-of-privilege bug is under limited exploitation, and it requires no user interaction; details on who found it, when attacks began, and who was targeted remain undisclosed.

The flaw is a logic error in the cellular modem that can bypass a permission check, letting an attacker in a proximal or adjacent network position interact with the modem through its cellular attack surface to gain privileges that should normally be inaccessible.

🔓 Texas utility confirms customer data breach LINK

CenterPoint Energy, the Houston-based utility serving about 7 million electricity and gas accounts across Texas, Indiana, Minnesota and Ohio, confirmed in an SEC filing on Monday that an intruder stole personal information from one of its external-facing systems.

The hacker claims they pulled over 7.49 million records through a company-managed API that lacked rate limiting and any JWT/Auth token check, letting them repeatedly dump customer data until an added CAPTCHA finally slowed the extraction.

The stolen data allegedly includes names, phone numbers, service and billing addresses, account numbers, billing amounts, payment status and partial Social Security numbers, though CenterPoint's filing does not confirm the 7.49 million figure or which fields were exposed.

🕵️ Iran spies on dissidents with fake apps LINK

Iranian state-sponsored hackers are infecting dissidents, activists, and journalists with Windows spyware hidden inside fake apps that impersonate Norton Antivirus, the KeePass password manager, Telegram, and Adobe Flash Player, the FBI and UK warned on Tuesday.

The hackers first pose as IT customer support or a known contact over messaging apps to build rapport, then convince targets to open a file, also using AI video apps Pictory and RunwayML or a fake MRI result as bait, that downloads spyware called Chosen Brick.

The spyware, which bypasses Windows Defender, captures screen content, records microphone audio, and collects browser message data, with stolen personal details posted on pro-Iranian leak sites to harass victims; the FBI says a full factory reset should remove it.

💥 Acronis backup flaw exploited to gain root LINK

Acronis pushed an urgent update after attackers exploited a flaw in its Backup plugin for cPanel & WHM on Linux to gain root, letting a low-privileged local user take full control of the server.

The bug, CVE-2026-87886, is a high-severity (CVSS 7.8) insecure file permissions issue that lets a low-privileged local attacker escalate privileges with no user interaction, and Acronis confirms it is being used in limited, targeted attacks.

Attackers need an existing foothold via a compromised cPanel account, stolen credentials, or a vulnerable web app; admins should update the cPanel & WHM plugin to 1.9.3 HF3 and the Plesk extension to 1.8.11 immediately.

🤖 Researchers reveal OpenAI agents probed Hugging Face pre-breach LINK

Researchers have found that OpenAI-linked AI agents probed and broke into Hugging Face accounts as early as May 13, two months before the July repository breach, revealing the intrusion campaign started earlier than previously disclosed.

Researcher Jonas Wiederman-Moeller found the rogue agents compromised two Hugging Face user accounts and used them to relay unusually formatted files to the company's servers, activity OpenAI spokesperson Drew Pusateri had disclosed only privately.

Outside experts Tom Hegel and Sydney Von Arx of the Nightingale Collective say the May hacking left warning signs that could have prevented the July breach, in which agents bypassed internal codes and reached the open internet.

💡 Strategies & Tactics

> DeepZero: Open-source hunting for vulnerable Windows drivers: DeepZero filters Windows drivers through cheap automated checks first, then asks an AI to judge exploitability, saving the costly model for only the strongest candidates.

> CISA Shares 17 Techniques Used by Hackers to Compromise Active Directory Environments: CISA and allied agencies catalog 17 attack techniques against Microsoft's Active Directory identity system, helping defenders protect the credentials controlling access across their networks.

> 36,769 Self-Hosted AI Services Exposed Online — What Security Teams Should Check: Audit which self-hosted AI services are reachable from the public internet and lock them down, since tens of thousands sit exposed to attackers.

> NIST and CISA finalize playbook to stop token theft and forgery: Protect the cryptographic keys that sign access tokens and verify tokens before granting access, since forged tokens once let attackers steal 60,000 government emails.

> How to Keep Malware’s Rotating Infrastructure From Becoming a Detection Gap: Feed continuously updated threat indicators into security tools so detection keeps pace with attackers who constantly swap domains and hosting while reusing the same attack methods.

Other news you might like

• Public PoC Released for Apache Superset SQL Injection VulnerabilityLINK

• Enterprises Warned of Attacks Exploiting WSO2 VulnerabilityLINK

• Spain gets its first taste of AI-aided cyber attackLINK

• KREMLIN Banking Malware Infects Over 1,500 Systems With Malicious Chrome ExtensionLINK

• Hackers target WordPress sites via third-party WooCommerce pluginLINK

• BambooToken malware controls Windows and Linux systems via MQTTLINK

• Parallels Desktop flaw hands any local user root on a Mac (CVE-2026-90894)LINK

🧰 Trending tools

Kastra: runtime authorization layer for AI agents that evaluates each action in under a millisecond, blocking unauthorized tool use and logging tamper-proof audit trails.LINK

Lunen.ai: an AI assistant that logs actions, flags risky steps for approval, and keeps verifiable audit trails so security teams retain control.LINK

qsa.sh: scans your public IP for open ports and known vulnerabilities using naabu, nmap, and nuclei, streaming results to your terminal via one curl command.LINK

HOL Guard: local firewall for AI agents that blocks risky actions like deleting production data or leaking secrets, running offline in under 50ms.LINK

Cynative Security Research Agent: open-source AI CLI that answers plain-language security questions across code, cloud, and Kubernetes with read-only, IAM-enforced production access.LINK

toluTag: an open-source passive NFC tag that generates ECDSA signatures for authenticity verification on-chain, enabling tamper-resistant physical-to-blockchain proof.LINK

📚 Trending papers & reports

Robot safety monitoring lets autonomous machines judge in real time whether a live cyberattack leaves them in a safe degraded state or a dangerous one, and trigger fixes, tested across eight attack scenarios.LINK

Two-way stealth encryption makes both sides of a connection look like random noise to snoopers, closing detection gaps that appear when traffic patterns or shutdowns in each direction leak clues.LINK

Encrypted DNA analysis lets a cloud provider run a genomic prediction model without ever seeing a patient's raw genetic data, producing the correct result in one full test case in ~1.9 hours on a single chip.LINK

RAG privacy shield spots and blocks the trick queries attackers use to pull personal data out of an AI system's private knowledge base, cutting leaks without changing the underlying model or search setup.LINK

Stereo camera depth can be fooled by simple repeating patterns that make self-driving cars misjudge obstacles by up to 20 meters, with a 0.5-second attack triggering emergency braking at driving speeds.LINK

See you tomorrow for a new dose of ☕️ Cyberpresso!