All tracked items across vulnerabilities, news, research, incidents, and regulatory updates.
This paper presents a new method for publishing data while protecting privacy using differential privacy (a technique that adds noise to data to hide individual information). The method is designed specifically for classification tasks (training AI models to categorize data), and it improves performance by keeping class labels (the correct categories for data records) unchanged and preserving correlations (relationships between different data attributes) using large language models and clustering algorithms.
Organizations are rapidly deploying AI agents (software systems that can perform tasks automatically) without securing the legacy infrastructure they depend on, creating a major security gap. Attackers can bypass AI-specific security measures by exploiting old vulnerabilities in underlying systems like unpatched servers, misconfigured permissions (Active Directory access controls), and cached credentials (stored login information), giving them access to the data and resources the AI agents use. The article demonstrates how this happens through a real attack example involving an S3 bucket (cloud storage), Lambda functions (serverless computing services), and overly broad access permissions.
Tencent, a major Chinese tech company, is testing an AI assistant called Xiaowei within WeChat (a messaging app with over 1.4 billion users), allowing users to interact via text or voice and access mini-programs (small apps that run inside WeChat). This move is part of Tencent's effort to compete with other AI companies in China's competitive market, though the company has not disclosed details about Xiaowei's capabilities or which AI models it uses.
CISOs (Chief Information Security Officers, the top security leaders in organizations) are expanding their roles beyond traditional IT security to also manage business risks, understanding how security decisions affect company revenue, operations, and strategic goals rather than just protecting systems and data. Security leaders are learning to master business risk by partnering with non-security teams like finance and legal, and by aligning their security strategies with the company's business objectives and key results (measurable goals that support company aims). This shift reflects a growing recognition that security and business risks are interconnected, requiring security leaders to advise executives on how their decisions impact the organization's ability to achieve its goals.
Samsung Electronics is deploying ChatGPT Enterprise and Codex (an AI tool that can write and review code) to all its employees in Korea and worldwide in its Device eXperience division, making this one of OpenAI's largest enterprise deployments. The company plans to use these tools across research, manufacturing, marketing, and other departments to improve employee productivity and problem-solving. ChatGPT Enterprise includes security features like data protection and access management so Samsung employees can use AI safely within company policies.
Cloudflare launched a feature that lets users deploy applications using Cloudflare Workers (a serverless computing platform) without creating an account by running a command that creates a temporary project lasting 60 minutes. The temporary deployment can be converted to a permanent project if the user claims it before the time expires.
vLLM (a system for running large language models) versions 0.10.2 through 0.12.x lack proper validation of sparse tensors (data structures with mostly empty values) when processing multimodal embeddings (numerical representations combining text and images). An attacker can send malicious embedding requests with invalid tensor indices to crash the system, exhaust resources, or potentially corrupt memory if the prompt-embeds feature is enabled.
vLLM versions 0.6.3 through 0.9.0 contain ReDoS (regular expression denial of service, where specially crafted text causes regex patterns to consume excessive CPU time) vulnerabilities in several components including the LoRA utility parser, phi4mini tool parser, and OpenAI chat endpoint. An attacker can send malicious input with nested or repeated structures to trigger severe CPU consumption and make the service unavailable.
North Korean hackers from the Sapphire Sleet group compromised an npm maintainer account (a person's credentials for publishing packages to npm, a JavaScript library repository) and used it to publish malicious updates to over 140 packages, injecting a fake dependency called "easy-day-js" that stole credentials, API keys, and cryptocurrency wallet information from developers' computers. The malware used a post-install hook (code that runs automatically when a package is installed) to download and execute additional harmful software, with different persistence techniques for Windows, Linux, and macOS systems.
Lloyds Banking Group is hiring 300 tech experts to work on agentic AI (autonomous artificial intelligence models that can plan and execute tasks with minimal human oversight) by September. While this hiring increases the bank's workforce now, the article notes that broader adoption of AI in the future could potentially lead to job cuts.
This research paper describes a new method for tracking users across multiple databases without exposing their private information. The method, called threshold multi-party private set intersection (T-MPSI, a technique where multiple parties can find matching data without revealing the full datasets to each other), improves the speed of existing approaches by 3 to 6 times, making privacy-preserving user tracking more practical for real-world deployment.
This research proposes a new method for detecting Out-of-Distribution data (inputs that differ significantly from the data an AI model was trained on) by using Kernel Principal Component Analysis, or KPCA (a technique that finds patterns in data using non-linear transformations). The method creates a special non-linear subspace (a mathematical space with special properties) from normal training data and checks if new inputs fit that pattern, while also addressing practical challenges like choosing the right mathematical kernel (transformation function) and handling large amounts of data efficiently.
Intelligence agencies from five countries (Australia, US, UK, New Zealand, and Canada) issued a joint warning that extremely powerful AI models capable of causing severe damage to governments and businesses could arrive within months, urging world leaders to take immediate action. The warning came after the Trump administration blocked foreign nationals from accessing Anthropic's Fable AI model, a highly anticipated AI system.
Patch the Planet is a program that uses AI models paired with human security experts to find and fix vulnerabilities in widely-used open-source software (software that anyone can view and modify). Instead of just reporting problems to already-busy maintainers, the program's security engineers validate issues, help develop patches, and create reusable tools so projects can continue improving security on their own.
Fix: The program provides participating projects with access to ChatGPT Pro, conditional access to Codex Security (an AI tool for code analysis), API credits for development and automation, and reusable security infrastructure including fuzzing harnesses (automated testing tools that feed random inputs to code), CVE analysis pipelines, differential-testing systems, threat models, expanded test suites, and workflows for deduplication, false-positive filtering, severity correction, and patch generation.
OpenAI BlogOpenAI is expanding Daybreak, a set of tools that uses AI models to help organizations discover and fix software vulnerabilities (security flaws in code) at scale. The initiative includes Codex Security (which automates finding and patching vulnerabilities), GPT-5.5-Cyber (a specialized AI model for cybersecurity tasks), and Patch the Planet (a program partnering with open-source projects to move from identifying vulnerabilities to fixing them).
Fix: OpenAI's Codex Security plugin 'implements what we've learned from internal and customer usage of our models into a solution to accelerate the process of discovering and patching vulnerabilities in existing systems as well as automatically preventing new vulnerabilities from ever reaching production.' Additionally, Daybreak provides tools to 'validate vulnerabilities, prioritize risk, generate and test fixes, and produce evidence inside existing security and development workflows.'
OpenAI BlogA cybersecurity tabletop exercise (a simulated attack-and-defense training scenario) showed how attackers might target a fictional supermarket's AI-powered supply chain system by exploiting weak credentials, poor network segmentation, and employee access to cause both operational damage and reputational harm through data leaks and disinformation. The attackers used stolen developer credentials, weak multi-factor authentication (a security method requiring multiple ways to verify identity), and phishing (fraudulent messages tricking people into revealing passwords) to breach systems, then deployed tactics like flooding the company with false security alerts, spreading deepfakes (AI-generated fake videos), and spreading misinformation on social media to amplify disruption.
Fix: The defenders established out-of-band communications channels (communication methods outside the normal network) to counteract the thousands of false alerts generated by attackers.
CSO OnlineAI agents are becoming workers inside organizations in Southeast Asia, while regulators enforce strict data residency rules and supply chain attacks threaten businesses. Zscaler proposes extending Zero Trust Exchange (a security model that verifies every access request, even from trusted sources) and SASE (Secure Access Service Edge, a cloud-based security platform) to protect AI agents alongside users and devices. The approach includes an AI Broker that inspects AI prompts and responses, endpoint security to find risky AI tools, and access mapping to track how AI uses data across cloud and on-premise systems.
Fix: The source explicitly describes Zscaler's technical solutions but does not frame them as fixes to a specific vulnerability or bug. Instead, they are presented as proactive security architecture recommendations. The key components mentioned are: (1) An AI Broker with an Agent Registry that inspects prompts and responses and enforces least-privilege access in real time; (2) Endpoint AI Security to expose risky local AI tools, browser extensions, and plugins; (3) An AI Access Graph and AI Protect that map AI assets, model usage, and data flows, backed by red-teaming, prompt hardening, and guardrails for more than 250 GenAI apps; and (4) Zero Trust B2B Exchange for policy-controlled partner access. N/A -- no patch, vulnerability fix, or incident mitigation is discussed in source.
CSO OnlineA reporter at The Atlantic discovered four publicly available datasets containing millions of songs (totaling between 100,000 and 12 million tracks each) that are being used to train AI models. These datasets have been downloaded thousands of times, and companies like Google and Stability have confirmed using them in their research, raising questions about how music is used in AI training without always crediting or compensating artists.
A speculative thought experiment called 'Europe 2031' imagines a future where Europe falls behind economically because the US and China invested heavily in AI datacenters (facilities housing the computer chips that power AI systems) and automation while Europe did not, leading to economic collapse and political instability. The scenario, which went viral among policymakers and EU officials, was created by Brussels-based thinktankers to warn Europe about the risks of falling behind in AI development and to highlight a communication gap between European policymakers and the US tech industry where most AI is being built.
SALT is a watermarking technique for diffusion models (AI systems that generate images by gradually removing noise from random data) that uses semantic guidance and adaptive latent space truncation to embed hidden ownership marks. The method aims to protect diffusion models from unauthorized use while maintaining the quality of generated images. This research addresses the need for better ownership verification and copyright protection in generative AI systems.