AI Sec Watch

Researchers developed R2BD, a new method to detect fake images created by AI (such as those made by diffusion models, VAEs, or GANs, which are different types of generative AI systems) by reconstructing them and measuring the differences from originals. The method is much faster (over 22 times speedier) than previous approaches because it works in a single step instead of many steps, and it works better across different types of AI image generators and different datasets.

Large text-to-image diffusion models (AI systems that generate images from written descriptions) can accidentally create inappropriate content like copyrighted artwork or offensive images, and existing removal methods often fail because unwanted concepts can reappear in subtle ways. The paper proposes SAFER, a method that identifies a concept-specific subspace (a mathematical region in the model's embedding space, which is how the AI represents meaning) associated with unwanted content and then projects prompts away from that region to remove the concept from generated images.

This research presents a new method for detecting deepfakes (AI-generated fake videos or images of people) that works better when the forgery techniques differ between training and testing. The approach uses orthogonal moment watermarking (embedding hidden marks in images using mathematical transforms), which stays intact when images are distorted but gets disrupted by deepfake manipulations, allowing the system to achieve 92.58% accuracy in detection.

Researchers discovered that deep neural networks (DNNs, a type of AI used for face and voice recognition) are vulnerable to embedding reconstruction attacks (ERAs, where attackers recover the original biometric data from the compressed numerical representation that the system stores). This attack works even when attackers have very limited access to the system, such as only seeing yes/no decisions or final scores, which is common in real-world commercial APIs.

OpenAI has launched Daybreak, an AI-powered cybersecurity platform that uses large language models (AI systems trained on vast amounts of text data) and agentic capabilities (the ability for AI to take independent actions toward goals) to help organizations find and fix software vulnerabilities faster. The platform competes with Anthropic's Claude Mythos and works through three stages: prioritizing threats, generating and testing patches in enterprise systems, and documenting results for verification. Daybreak is being rolled out across three versions of GPT-5.5, from general-purpose use to specialized cybersecurity workflows.

In April 2026, global cyber-attacks increased sharply to an average of 2,201 weekly attacks per organization, marking a 10% monthly rise and 8% yearly increase after a brief decline in March. Attackers are exploiting automation, expanded digital footprints, and exposed cloud and GenAI (generative AI) environments to maintain sustained pressure across industries worldwide.

Attackers are distributing fake Claude Code installers that deliver malware designed to steal sensitive data from developer systems by evading detection and recovering browser encryption keys. The malware uses a PowerShell loader (a script-based delivery method) to hide malicious activities and exploits Chrome Elevation Services to bypass Application-Bound Encryption (ABE, a Chrome protection added in version 127 to prevent password and cookie theft).

Hundreds of software packages on npm (Node Package Manager) and PyPI (Python Package Index) were compromised in the Shai-Hulud attack campaign, which used stolen OIDC tokens (authentication credentials that verify a developer's identity) to publish malicious versions with valid cryptographic signatures, making them appear legitimate. The malware targets developer credentials like GitHub tokens, AWS secrets, and SSH keys, then hides itself in code editor auto-run tasks so uninstalling the packages doesn't remove it. The attack affected popular projects including TanStack, Mistral AI, Bitwarden, and others.

Testing Anthropic's Claude Mythos AI model on the curl data transfer tool found only one actual low-severity vulnerability in 178,000 lines of code, despite Anthropic's claims that the model could identify thousands of zero-day vulnerabilities (previously unknown security flaws). Experts are divided on whether this result shows that Mythos is less powerful than claimed or simply that curl's code is already very secure from previous audits and analysis by other tools.