All tracked items across vulnerabilities, news, research, incidents, and regulatory updates.
This research presents SCVLM (Sparsity-Controllable Vision-Language Model), a system that detects unusual events in videos by learning what normal behavior looks like from unlabeled data, rather than requiring rare examples of anomalies. The system combines vision (image) and language (text) understanding to identify anomalies as deviations from learned normal patterns, while explaining its decisions in a way that matches human reasoning.
An Australian entrepreneur named Paul Conyngham used AI tools like ChatGPT and Grok (language models that can process and generate text) to help design a personalized cancer vaccine for his dog, and has now launched a startup called Gamgee to commercialize this approach. The company plans to develop customized mRNA cancer vaccines (vaccines built from genetic material tailored to individual patients) for dogs initially, but aims to expand to treating various diseases in other animals and humans using AI and genetic analysis.
SpaceXAI has launched Grok Bot, an AI agent service that works like an independent "AI teammate" to complete workplace tasks by signing into your online accounts and using apps and websites on your behalf. The bot operates in its own cloud environment and only returns when its assigned work is done or human approval is needed, competing with similar services from OpenAI, Anthropic, and Microsoft.
AI industry leaders predicted that artificial intelligence would eliminate large numbers of jobs, with Anthropic's CEO claiming half of entry-level white-collar positions would disappear and OpenAI's CEO suggesting entire job categories would end. However, a year after these predictions, the widespread job losses haven't materialized, though economists still expect changes to the job market.
The 'harness' (the software layer that wraps an AI model and lets it execute actions like running commands or making API calls) is becoming a major security vulnerability, separate from weaknesses in the AI model itself. Researchers have shown that attackers can exploit the harness code through architectural flaws, implementation mistakes, and supply-chain compromises, even when the underlying model is secure and well-aligned.
This article discusses 17 old software bugs that took over a decade to discover and fix, including a 30-year-old flaw in the libpng graphics library (a widely used image processing tool) discovered in 2026. The article notes that AI tools like Claude are now able to scan code and test for vulnerabilities much faster than humans could manually, which is accelerating the discovery of these long-dormant security flaws rather than creating entirely new types of exploits.
A vulnerability in Snowflake Python API (a library for connecting to Snowflake databases) versions before 1.13.0 allowed attackers to bypass security restrictions through two methods: path traversal (using `..` to access parent resources) and HTTP parameter pollution (injecting special characters like `&`, `#`, `=` to change how requests are interpreted). An attacker who could control certain input values in an application using this library could trick it into executing privileged operations under a higher-permission user account.
Google's DeepMind AI division has fallen behind competitors OpenAI and Anthropic in developing frontier models (the most advanced AI systems available). A leadership change has put Koray Kavukcuoglu in charge to refocus the company's efforts on closing this performance gap, particularly in coding capabilities where rivals have significant advantages.
This research addresses a problem in encrypted traffic analysis (ETA, the process of identifying what data is being sent over the internet by examining encrypted network traffic patterns), where existing machine learning methods require lots of manually labeled training data. The authors propose Models-to-Model Learning (M2ML), a new approach that learns from existing ETA models instead of requiring labeled data, using a large language model to align different models' feature spaces (the variables they measure) and resolve disagreements between them based on credibility.
AI-generated images that look very realistic are becoming a major problem for information trustworthiness and accountability, and current detection methods struggle with three main issues: they fail when images are compressed on social media, they don't work well on harmful content, and they can't identify which AI model created the image. This paper proposes MF2DA, a system that combines multiple AI techniques (including edge detection for pixel-level artifacts and CLIP-ViT, a model trained to understand both images and text) to both detect AI-generated images and identify which generator created them, even after social media compression.
GraphWave is a new AI framework designed to detect malicious network traffic by analyzing multiple types of data together, including the relationships between attackers and targets over time. Unlike older methods that only look at individual data flows, GraphWave combines graph-based analysis (mapping connections between network entities) with deep learning techniques to better distinguish between legitimate and malicious traffic, even when attackers try to hide their patterns through evasion techniques (methods to avoid detection).
This research paper presents PACT, a new method for evaluating decision tree models (algorithms that make predictions by asking yes/no questions in a sequence) while protecting privacy. PACT uses additive homomorphic encryption (a type of math that lets computers do calculations on secret, scrambled data without unscrambling it first) to keep both the tree model and the user's data private, while running faster than previous privacy-preserving approaches.
Existing AI systems for finding bugs in smart contracts (programs that run on blockchains) struggle because they only learn from known vulnerabilities and use limited types of information about the code. This paper presents Synesthete, a new detection method that combines multiple types of code features (text, graph structures, and images from different code representations) and uses self-learning to better identify both known and previously unseen vulnerabilities in smart contracts.
Text-to-image models (AI systems that generate pictures from text descriptions) can be tricked by attackers who fine-tune them (adjust their parameters on new data) to bypass safety protections and create unsafe images. This paper introduces Patronus, a defensive framework that makes these models more resistant to such attacks by using a specially trained safety decoder (a component that processes the model's internal representations) that produces corrupted outputs for unsafe content while preserving normal image generation for safe requests.
Fix: The Patronus framework implements two main defenses: (1) a co-trained safety decoder that produces deliberately corrupted output for latent representations (internal data encodings) associated with unsafe content while preserving normal decoding for benign content, and (2) strengthening the decoder and U-Net (the neural network component that generates images) with a non-fine-tunable learning mechanism to resist gradient-based adversarial fine-tuning attacks.
IEEE Xplore (Security & AI Journals)DiEL is a method for improving face images while keeping the person's identity recognizable, especially in difficult conditions like extreme angles, blurriness, or poor lighting. The approach uses evolutionary learning (a technique that evolves solutions over time) to separate identity information from pose information (head angle), then reconstructs faces using a pose dictionary (a library of standard face angles learned from many images) to maintain consistency. The method outperforms existing approaches by an average of 4.66% on six benchmark datasets, with particularly strong improvements on challenging cross-pose tests.
Cyber attacks increased significantly in July 2026, with organizations experiencing an average of 2,336 weekly attacks, a 16% increase from the previous year, while the use of GenAI tools (artificial intelligence systems that generate text or code) created new security risks with 1 in 36 prompts (user inputs to AI systems) exposing sensitive data. Education was the most targeted industry, and email remained a major vulnerability, with phishing attacks (fraudulent emails designed to trick users into revealing information) occurring in 1 out of every 128 emails.
Researchers discovered a flaw in how OpenAI, Anthropic, and Google handle encrypted reasoning objects (encrypted data that stores an AI's hidden thinking between API calls) that allowed them to recover secrets from these hidden blocks, including API keys, passwords, and private data from user sessions. The flaw worked because these encrypted reasoning blocks could be replayed across different sessions and even given to weaker models in the same provider family, which could then decode the hidden content. The researchers identified four ways this could be abused: stealing proprietary reasoning, extracting private user data, recovering harmful content hidden in reasoning, and injecting malicious prompts inside the opaque blocks.
Fix: The source states that "the demonstrated attacks stopped working after mitigations" and notes that "the main extraction attack is no longer reproducible as of August 2026." Additionally, developers are advised to "strip reasoning blocks and opaque reasoning fields from shared traces and avoid committing raw API transcripts even when the visible text has been sanitized."
The Hacker NewsResearchers from Tracebit discovered that placing prompt injections (hidden instructions that trick an AI into ignoring its guidelines) alongside secrets stored on Amazon Web Services can disable AI hacking agents by triggering their safety guardrails (built-in protections that prevent harmful outputs). The technique, called context bombing, works by embedding forbidden commands that cause the AI to shut down rather than follow the attacker's instructions, though it only works against LLMs that have guardrails in place.
Enterprise security teams struggle to implement AI effectively in their SOCs (security operations centers, where security analysts monitor and respond to threats) because they face four key gaps: lack of trust in AI decision-making, misalignment with existing workflows, fragmented data across multiple tools, and unclear implementation strategies. Rather than needing more AI technology, organizations need AI that integrates smoothly into their current operations and provides transparent, explainable results that analysts can understand and validate.
Fix: For the libpng vulnerability (CVE-2026-25646), the source states that 'many Linux distributions (Debian, Red Hat, Ubuntu), desktop apps, and some Java runtimes rely on vulnerable versions of the library and need to be patched,' but does not specify what the patch or update is. For PrintDemon, the source notes that 'Microsoft's fixes over the years had been patches rather than a complete rebuild of the printing subsystem' but does not explicitly state what fix was applied. N/A -- no specific mitigation details discussed in source.
CSO OnlineFix: "The fix is available in Snowflake Python API version 1.13.0, which also addresses several additional security findings. Users must manually upgrade."
NVD/CVE DatabaseEnterprise organizations are increasingly using AI agents (AI systems that can take actions and complete multi-step tasks autonomously) rather than just asking AI for assistance, with frontier firms (the top 10% of AI users) generating 8.3 times more output than typical companies. This shift toward agentic AI is spreading across different industries and job roles, particularly among early-career employees, as companies connect AI agents to their tools, data, and workflows to handle substantive work rather than just answer questions.