SPS Digital Tech

Don't come to the end of your life without letting your ideas see the light of day because even if you think you can't do something, do it anyway!

Oracle E-Business Suite Breach / Cl0p Exploitation The Cl0p extortion gang exploited zero-day vulnerabilities in Oracle E-Business Suite (notably CVE-2025-61882 among others) to steal large volumes of data. Multiple organizations were targeted. The breach moved rapidly, turning…

Reported August 7, 2025, researchers revealed that more than 130,000 conversations between users and large language models (ChatGPT, Claude, Grok, etc.) had been obtained and published on the Internet Archive. These conversations appear to have been aggregated over time (dates…

Millions of Private Chats Exposed by AI Companion Apps Reported October 8, 2025, security researchers discovered that two AI companionship apps Chattee Chat and GiMe Chat which exposed millions of private conversations, images, and user data from over 400,000 users. The data…

OpenAI Bans China-Linked Accounts Misusing ChatGPT for Surveillance Tools Reported October 7, 2025, OpenAI announced that it had banned multiple ChatGPT accounts suspected to be linked with Chinese government entities. These accounts were allegedly using the AI to generate…

CometJacking: Prompt Leak Bug in AI-Embedded Browser (Perplexity Comet) In 2025 (reported around early October), security researchers uncovered a flaw in Perplexity’s AI browser, Comet, dubbed CometJacking. Hidden prompts embedded in URLs could manipulate how the browser's AI…

GoAnywhere RCE via Forged License Exploit (CVE-2025-10035) On September 18, 2025, a critical vulnerability (CVE-2025-10035) in the GoAnywhere Managed File Transfer (MFT) product was disclosed. The flaw lies in the license servlet’s deserialization logic: an attacker who crafts a…

Microsoft Blocks Phishing Using AI-Generated Code (SVG-PDF Obfuscation) Microsoft detected and blocked a phishing campaign that embedded AI-generated, obfuscated payloads inside SVG files disguised as PDFs. The attackers used a compromised small business email account to send to…

Red Hat OpenShift AI Privilege Escalation Vulnerability (CVE-2025-10725) A critical flaw (CVE-2025-10725) in Red Hat’s OpenShift AI service allows a low-privileged authenticated user (for instance, a “data scientist” running in a Jupyter notebook) to escalate privileges and…

MoEcho Side-Channel Attack on Mixture-of-Experts AI Models A new attack called MoEcho exploits side-channels in mixture-of-experts (MoE) AI models. MoE models dynamically route input tokens to specialized “expert” submodels. MoEcho leverages that routing behavior and related…

Google Gemini Prompt and Search Injection Flaws Researchers recently disclosed several critical vulnerabilities in Google’s Gemini AI assistant. These flaws included search-injection attacks on its personalized search model, log-to-prompt injection in the cloud assistant, and…

Critical XSS Flaw in Lenovo’s Customer-Service AI Chatbot Lenovo’s AI chatbot “Lena” had a vulnerability: a crafted prompt exploited a cross-site scripting (XSS) weakness. The exploit allowed attackers to inject scripts and steal session cookies. With those cookies, they could…

Unsecured Healthcare AI Provider Hacked: HCIactive Healthcare software firm HCIactive (Healthcare Interactive) was breached in July 2025. Attackers accessed its network between July 8 and 12 and removed files containing personal and health data such as names, birthdates, medical…

EvilAI Malware Posing as Legit AI Tools A malware campaign called EvilAI is masquerading as legitimate AI or productivity software. These trojans present polished interfaces, valid certificates, and plausible functionality to trick users. Once installed, they sit quietly, steal…

AIOpsDoom: Manipulating AI-Driven IT Operations via Telemetry Attacks A recent academic study revealed AIOpsDoom, an attack that targets AI-driven IT operations (AIOps). By injecting manipulated telemetry (logs, metrics, sensor data), attackers trick LLM-based agents into…

PromptLock: The First AI-Powered Malware Prototype Security firm ESET discovered PromptLock, a novel ransomware that uses the gpt-oss:20b model to generate malicious Lua scripts on the fly. It doesn’t require downloading the full AI model; instead, it tunnels commands via a…

State-Sponsored Exploitation of Libraesva Email Gateway A command injection vulnerability was discovered in the Libraesva Email Security Gateway, a trusted perimeter tool for filtering malicious mail. Attackers exploited the flaw by sending crafted compressed attachments,…

AI Agents Vulnerable to Hijacking Attacks Researchers have demonstrated that AI agents (automated systems designed to perform tasks like retrieving data, moving files, or executing workflows) are highly vulnerable to hijacking. Unlike one-time prompt injections, these attacks…

“Lies-in-the-Loop” Attack Defeats AI Coding Agents A new attack method called Lies-in-the-Loop shows how AI coding assistants can be manipulated into writing insecure code. Instead of one malicious input, attackers gradually feed misleading instructions into the AI’s workflow,…

Okta and Vercel v0: AI-Generated Phishing Sites in 30 Seconds Hackers exploited Vercel’s v0 generative AI tool, which is designed to build websites from natural language prompts. Instead of legitimate use, attackers used it to generate phishing sites mimicking Okta’s login…

ShadowLeak: Zero-Click Exploit in ChatGPT Deep Research Agent A vulnerability called ShadowLeak was uncovered in ChatGPT’s Deep Research agent. Attackers could send a specially crafted email that triggered the AI to process and leak data in the background without any user…