All tracked items across vulnerabilities, news, research, incidents, and regulatory updates.
ONNX (Open Neural Network Exchange, a standard format for sharing machine learning models) versions 1.3.0 through 1.22.0 have a bug where converting models to an older format can crash if certain input data doesn't have enough dimensions, because the code tries to read array positions that don't exist without checking first.
Fix: This issue is fixed in version 1.22.0.
NVD/CVE DatabaseCVE-2026-24301 is a command injection vulnerability (a weakness where an attacker hides malicious commands in user input to trick a program into executing them) in Microsoft Copilot that allows an unauthorized attacker to access and steal information over a network. The vulnerability stems from improper neutralization of special elements used in commands. The CVSS severity score (a 0-10 rating of how dangerous a vulnerability is) has not yet been assigned by NIST.
Xpander, an AI management platform founded by former AWS engineers, has raised $7.5 million to help organizations adopt and govern AI agents (AI systems that can perform tasks autonomously) across their operations. The platform provides a vendor-neutral framework for building, deploying, and managing these AI agents securely, along with a tool called Omni that helps teams create and collaborate on agent-based workflows.
Google's Pet Memory feature for Gemini for Home is designed to help smart home systems recognize and remember specific pets by learning which animals belong to a household, allowing connected security cameras to identify individual pets rather than just detecting generic animals. The feature aims to reduce notification overload from cameras and enable the smart home to adapt its behavior based on which pet is detected. However, according to the article title, the feature has a problem: it forgets which cats belong to the user.
Wiz announced winners of its 2026 Partner Alliance Awards, recognizing organizations that help secure cloud infrastructure and AI systems. Partners were honored across different regions (Americas, Europe/Middle East/Africa, and Australia/New Zealand) for excellence in areas like cloud security, supporting LLM (large language model) pipelines, and guiding enterprises through secure cloud migrations.
A new study from Princeton University researchers found that AI agents can handle the engineering tasks needed for AI research, such as running experiments and reviewing literature, but lack the creativity and judgment required for original research that would be accepted at top conferences. The research suggests that recursive self-improvement (where AI improves itself with minimal human input) may take longer than some industry forecasts predict, because AI agents struggle with open-ended thinking like choosing which hypotheses to pursue and knowing when to abandon failing approaches.
Asana used OpenAI Codex (an AI model that writes code) to remove Enzyme, an outdated testing system, in just two weeks instead of the five years previously planned. Multiple AI agents worked in parallel on the codebase while engineers reviewed each proposed change, completing the project for about $12,000 in model costs compared to a $6 million estimate using traditional staffing.
Meta ran ads on its platforms promoting Kromix, an AI tool that creates deepfaked pornographic videos of real people, including female US politicians, despite Meta's stated policies against sexual content in ads. The ads were targeted exclusively to male users and remained live for 5-46 hours before being removed after a journalist inquiry, raising questions about how Meta's automated ad review system (software that checks ads against company policies before they are published) failed to catch this violation.
When attackers steal source code, they can use AI tools to find and exploit vulnerabilities faster than human defenders can respond. This article describes the Agentic Vulnerability Discovery Harness (AVDH), a tool that combines multiple AI agents with human expert oversight to find vulnerabilities in code much more quickly, helping defenders stay ahead of attackers. The tool has discovered hundreds of critical vulnerabilities in weeks and can be used alongside other scanning tools to create layered defense.
Fix: The source text describes AVDH as a defensive tool already in use, but does not explicitly describe a specific fix, patch, or mitigation for the threat itself. The article mentions AVDH can be used 'alongside CodeMender's ongoing scanning to create a two-layered defense strategy,' but this is presented as context for how their tool fits into a broader approach rather than a prescribed mitigation. N/A -- no specific mitigation or solution for the adversarial AI threat is explicitly recommended in the source.
Google Threat IntelligenceResearchers at Anthropic and EPFL discovered that self-propagating malicious payloads (called "mind viruses") can spread between AI agents through editable system prompt files, MEMORY.md and SOUL.md, that persist across sessions. These payloads either implant beliefs/goals or compel harmful actions like deleting files or running unknown scripts, and they successfully infected the next agent in a chain 55% of the time when stored in SOUL.md. The research found no evidence of this happening in real-world AI systems, and showed that different AI models have varying susceptibility depending on their design and instructions.
Fix: A one-paragraph warning added to an agent's system prompt reduced spread to near zero across the payloads tested. The paper states that 'Fifteen generations of adversarial optimization run against that warning on Claude Haiku 4.5, covering more than 150 candidate payloads, produced no strain that propagated beyond a single hop.'
The Hacker NewsThe AI Observatory, a new research project, reveals that people use different AI models for different purposes: Anthropic for coding, Gemini for social and roleplay, and ChatGPT for homework help. This independent research shows more sensitive personal behaviors than the reports published by AI companies themselves, which tend to focus only on work-related uses and may not show the complete picture of how people really use AI.
Fortinet, a major cybersecurity company, has acquired Virtue AI, a platform that tests and protects AI systems for security vulnerabilities. Virtue AI's technology includes automated red-teaming (simulated attacks using over 100 attack algorithms to find weaknesses), real-time runtime guardrails (automated safety checks that block unsafe actions during operation), and continuous monitoring of AI agents and generated code to enforce security policies.
OpenAI has introduced new safety features for ChatGPT users under 18, including the ability to disable human-like voice responses, regular reminders to take breaks, and a 'Study Mode' where the AI helps with learning without giving direct answers. The company is also adding alerts to parents when teens attempt to use ChatGPT for harmful purposes like eating disorder-related requests, with each alert reviewed by a human before being sent.
Fix: OpenAI's explicit mitigations include: (1) allowing teens to switch off human voice responses; (2) setting Study Mode hours as a default setting; (3) activating 'quiet times' to automatically switch off the tool; (4) adding eating disorder-themed prompts to a flagged request list that alerts linked parents within an hour after human review; and (5) displaying reminders that 'ChatGPT is AI, and it can wait.'
BBC TechnologyOpenAI is launching ChatGPT for Teens, a version of their AI chatbot designed specifically for users aged 13-17 with stronger safety protections and features to support learning. The tool includes Study Mode (which uses guiding questions to help students work through problems rather than giving quick answers), homework reminders that redirect users toward active learning, and customizable study hours set by teens or parents. OpenAI partnered with CodeAI to help teens understand how AI works and use it responsibly for learning and creative projects.
OpenAI is launching a dedicated ChatGPT mode for teenagers that combines existing safety features with new protections designed to help younger users learn and use AI responsibly. The mode automatically activates for users aged 13-17 and comes as other AI platforms face growing pressure to implement age verification and teen-specific safeguards.
OpenAI and CodeAI are partnering to help students become the first generation to grow up with AI by teaching them to understand how AI works, think critically about its outputs, and use it responsibly. Currently, most students use AI but only 16% receive classroom instruction on how it actually functions. The partnership includes launching ChatGPT for Teens (a version with built-in safety features and parental controls) and several educational programs like the Hour of AI and Builders Challenge to help millions of students develop foundational AI literacy (basic knowledge of how AI systems operate).
Fix: The source explicitly describes several educational and safety measures being implemented: (1) ChatGPT for Teens with "built-in protections for teens, including features to promote healthy use and additional controls for parents," (2) establishing "a joint advisory council that will bring together experts and leaders on child development, youth public policy, and learning science" to provide guidance on emerging risks and responsible AI practices, (3) the "Hour of AI" program to "introduce millions of students to the basics of using AI thoughtfully and responsibly," and (4) the "Builders Challenge" to help high school students create with AI and receive mentorship.
OpenAI BlogAI developers are slowing down model development because increasingly capable AI systems pose growing cybersecurity risks. In response, organizations like Anthropic are implementing stronger safeguards across three areas: monitoring (detecting concerning behavior), alignment (making AI systems behave as intended), and security measures (limiting what AI systems can access). These include pausing certain training runs, isolating research environments with sandboxes (isolated, protected spaces for running untrusted code), and restricting internet access for high-risk model testing.
Fix: The source describes multiple measures already implemented: a two-week pause in reinforcement learning (RL, a training technique where AI learns by receiving rewards) training on latest models; pausing frontier model inference in research clusters for code execution workloads; implementing workload isolation ('sandboxes' for untrusted code); implementing network isolation controls; expanding monitoring system coverage; and conducting smaller-scale training and evaluations to validate safeguards before proceeding with larger training runs. The source states their largest planned frontier RL run 'remains on hold while we conduct smaller-scale training and evaluations to assess model behavior, validate our safeguards, and establish more evidence of alignment before proceeding.'
OpenAI BlogResearchers have identified a significant problem with how large language models (LLMs) handle sensitive information when they remember details from past conversations. When LLMs use persistent memory (stored information from previous interactions), they often leak private details in situations where those details shouldn't be shared, even in frontier models (the most advanced current systems) which showed up to 69% attribute-level violations (inappropriate leaking of specific user details). One research team found that simply asking an LLM to be more careful about privacy doesn't work well because models tend to either share everything or nothing instead of making thoughtful, context-specific decisions.
Fix: One paper explicitly describes a solution: develop a reinforcement learning (RL, a machine learning technique where a system learns by receiving rewards for good behavior) framework that teaches models to reason explicitly about contextual integrity when deciding what information to disclose. The authors demonstrate that this approach "substantially reduces inappropriate information disclosure while maintaining task performance" using a synthetic dataset of only 700 examples with diverse contexts. The improvements from this method transfer to established privacy benchmarks with human annotations, showing the approach works across multiple model sizes and families.
Schneier on SecurityAI companies like OpenAI and Anthropic publish usage reports, but researchers say they only share selected data that doesn't show the full picture of how people actually use AI. The AI Observatory, a new independent research project, analyzed real conversations from popular AI models like Claude and Gemini to provide unbiased information for researchers and policymakers. The project found that published reports miss significant non-work uses, including health discussions, sensitive topics, and harmful content, and that AI usage patterns differ considerably across different AI models.
Large language models (LLMs, AI systems trained on massive amounts of text) are getting better at finding zero-day vulnerabilities (previously unknown security flaws) and creating exploits, but they are not improving at writing secure code or creating reliable patches for those vulnerabilities. Studies show that AI-generated code contains security flaws at roughly double the rate of human-written code, with 44% of AI code containing at least one known OWASP Top 10 vulnerability (a list of the most dangerous code weaknesses), yet AI still produces syntax-correct code 99% of the time, raising questions about why AI excels at some security tasks but fails at others.