All tracked items across vulnerabilities, news, research, incidents, and regulatory updates.
Google now allows users to remove visible watermarks from AI-generated images, videos, and music in Gemini and its Flow video generator by toggling off a 'Media watermark' setting. While the visible "sparkle" watermark in the bottom-right corner can be disabled, Google still embeds invisible SynthID watermarks and C2PA metadata (hidden identification markers that track AI-generated content) in the background of all generated content.
Fix: To remove visible watermarks, toggle off the 'Media watermark' setting in Gemini or Google's Flow video generator.
The Verge (AI)Oracle has released Database Security Central, a free tool available until February 2027 that gives organizations a single place to monitor security risks across their databases. The tool helps security teams check security health, spot configuration drift (when actual settings differ from intended settings), identify access risks, track sensitive data, and manage security policies centrally. This comes as Oracle faces increased attacks on its database products and the emergence of Mythos, a bug-hunting AI model (an AI designed to find software vulnerabilities).
This newsletter covers multiple business and tech topics, including Anthropic's early investor meetings for a potential IPO (initial public offering, a company's first sale of stock to the public), OpenAI's revenue chief departing, and Apple opening a new manufacturing facility in Houston. The article does not describe any AI-related security issues, bugs, or vulnerabilities.
This essay argues that OpenAI and Anthropic, founded to develop AI safely in the public interest, have become corporate companies focused on investor profits rather than public benefit. As these companies face declining stock valuations and questions about long-term profitability, the essay proposes that if they fail financially, the US government should nationalize them and operate them as public research labs under democratic control.
Apple has partnered with Alibaba to develop a custom large language model (LLM, a type of AI trained on large amounts of text data) specifically for the Chinese market, marking a shift from its previous approach. This collaboration gives Apple greater control over its AI products in China's competitive smartphone industry, though it reflects the complex relationship between U.S. and Chinese tech companies.
Trigger.dev, a platform for building AI agents and workflows, had a security vulnerability from version 4.4.2 through 4.5.0-rc.5 where user input was improperly validated when creating file paths. An attacker with a valid API key could exploit path normalization (the process of simplifying file path references like converting '../' to parent directories) to access or modify files belonging to other customers' data. This vulnerability allowed unauthorized access to task payloads (the data that workflows process).
Trigger.dev, a platform for building AI agents and workflows, had a security flaw in versions 4.4.2 through 4.5.0-rc.4 where an API endpoint didn't properly check which environment a user belonged to before replaying task runs (a process that re-executes a previous job). This allowed someone with a valid API key to replay another organization's tasks, waste their resources, and potentially execute malicious code by overwriting task data through a separate vulnerability.
Microsoft is removing Mico, an animated avatar character (a yellow blob that reacts with facial expressions) from Copilot's voice mode and moving it to Learn Live, a Microsoft educational platform. Mico was introduced last October to give Copilot a visual identity, but will now be used in a different context where it has "more to react to."
A 17-year-old from Massachusetts was arrested and accused of killing his mother and brother, with prosecutors investigating whether his use of ChatGPT (a conversational AI system) played a role in the crimes. Authorities report he used the internet and AI to search for fantasy stories about killing family members.
IBM Langflow OSS (open-source software) versions 1.0.0 through 1.9.6 has a security vulnerability where an attacker can gain unauthorized access to user accounts because the system doesn't properly limit how many times someone can try to log in with wrong credentials. This weakness, called CWE-307 (improper restriction of excessive authentication attempts), allows attackers to keep trying passwords without being stopped.
Trigger.dev, a platform for building AI agents and workflows, had a security flaw in versions before 4.5.6 where an API endpoint didn't properly check which project owned a deployment. This allowed someone with a valid API key for one project to hijack another project's deployment by attaching their own background worker (a component that runs tasks in the background) to it and changing its status.
Trigger.dev, a platform for building AI agents and workflows, had a security flaw in versions before 4.5.2 where the Google authentication function didn't check if a user's email was verified by Google. This allowed an attacker to create a Google account with an unverified email matching someone else's account, then use it to take over that existing account.
Trigger.dev, a platform for building AI agents and workflows, had a security flaw in versions 3.3.8 to 4.5.6 where the PUT /api/v1/runs/:runId/metadata endpoint (a web address for updating run information) accepted attacker-controlled input without proper filtering. This allowed attackers with a normal API key to perform prototype pollution (a type of attack that corrupts shared object properties in JavaScript), which could break database queries, disrupt other users' authentication, and crash the application.
Kibana (a data visualization tool) has a missing authorization bug where its Machine Learning feature doesn't properly filter data between spaces (isolated work areas). This means operations from one space could access and modify machine learning data from all other spaces in the system, causing unauthorized information disclosure and data changes.
A security flaw in Kibana (Elastic's data visualization tool) Machine Learning allows users to remove trained models (pre-built AI models) from a workspace if they have permission to create certain types of jobs, even if they shouldn't have that permission. The actual model isn't deleted and can be restored by someone with proper access, but this is still a privilege escalation vulnerability (a situation where someone gains more control than they should have).
Elasticsearch has a vulnerability in its native inference process (the system that runs uploaded machine learning models) where it doesn't check that memory addresses (locations in computer memory) stay within allowed boundaries, allowing a user with model upload privileges to read or write memory outside intended areas, potentially crashing the process or executing arbitrary code.
The llm-gemini 0.33 plugin update adds support for newer Google Gemini AI models (including Gemini 3.7 Flash and embedding models, which are AI systems that convert text into numerical representations) and improves compatibility with LLM 0.32, enabling features like reasoning traces (showing an AI's step-by-step thinking) and server-side tools. However, the update introduces a browser compatibility issue where Firefox and Chrome fail to display SVG (scalable vector graphics, a format for creating images) elements correctly due to stricter standards, while Safari renders them properly.
This paper presents REI-Net, a new AI network that quickly extracts internal fingerprints (fingerprint patterns beneath skin) and internal sweat pores from OCT (optical coherence tomography, a medical imaging technique that captures 3D images below the surface). The method is much faster than previous approaches, completing analysis in 605 milliseconds, and achieves better accuracy for biometric security applications.
Fix: Oracle accelerated its patching schedule, switching from quarterly releases to monthly releases. The first monthly batch fixed 35 flaws.
CSO OnlineCyera acquired Oasis for $1 billion to combine data security (protecting sensitive information) and identity management (controlling who can access systems) into one unified control system for AI agents (software programs that act autonomously). The new system will grant access based on business needs and context rather than fixed permission roles.
At Black Hat USA 2026, security experts highlighted that AI is making it easier for attackers to find and exploit vulnerabilities, so traditional monthly patching is no longer sufficient. New risks include trojanized AI skills (instruction files for AI agents) being distributed through software marketplaces and supply-chain attacks using forged commits and token misuse on platforms like GitHub. The most effective AI-powered security research combines human expertise with AI capabilities rather than letting AI work autonomously.
Fix: GitHub Threat Detector, an open-source tool released by Microsoft researchers Yossi Weizman and Mor Weinberger, offers 30 built-in detection rules to identify supply-chain attacks on GitHub by analyzing GitHub webhooks, APIs, and Git metadata for suspicious patterns like forged commits and workflow abuse. Additionally, organizations should adopt memory-safe languages such as Rust, use AI-assisted engineering to improve existing codebases, and automate remediation (fixing issues automatically) rather than relying on monthly patch cycles.
CSO OnlineFix: This issue is fixed in version 4.5.0-rc.5. Users should upgrade to this version or later.
NVD/CVE DatabaseFix: This issue is fixed in version 4.5.0-rc.4.
NVD/CVE DatabaseFix: This issue is fixed in version 4.5.6.
NVD/CVE DatabaseFix: Update to version 4.5.2, which fixes this issue.
NVD/CVE DatabaseFix: Update to version 4.5.6 or later.
NVD/CVE Database