AI Sec Watch

This research addresses leaky private information retrieval (L-PIR), a system where a user queries a database while accepting some measurable privacy leakage, quantified using differential privacy (a mathematical framework that bounds how much information an observer can learn about individual data). The researchers developed an improved L-PIR scheme that reduces privacy leakage from O(K) to O(log K) by using cyclic permutations (mathematical rearrangements) and assigning higher probabilities to keys with lower Hamming weights (fewer 1-bits in binary representation), achieving better privacy-to-cost tradeoffs than previous methods.

Researchers have developed a new backdoor attack method called Trigger without Trace (TwT) that can secretly compromise text-to-image diffusion models (AI systems that generate images from text descriptions) while avoiding detection. The method works by using syntactic structures (grammar patterns) as hidden triggers and employing a mathematical technique called Kernel Maximum Mean Discrepancy (KMMD, a way to match statistical distributions) to make malicious samples look identical to legitimate ones, achieving a 97.5% success rate while bypassing three existing defense detection systems.

Vul-CTG is a new AI framework for detecting software vulnerabilities (bugs that create security weaknesses) by combining two approaches: PLMs (pretrained language models, AI systems trained on large amounts of text) and GNNs (graph neural networks, AI systems that analyze connected data structures). The framework improves on existing methods by better combining code text analysis with program graph analysis, using contrastive learning (training the AI to recognize similarities and differences) and handling unreliable training labels, achieving about 3% better accuracy than previous approaches.

Model extraction attacks (MEA, where attackers steal the functionality of AI models by creating a clone with similar behavior) are a security threat that defenders counter by using auxiliary data to make the victim model give misleading predictions. However, realistic auxiliary data is hard to obtain, gives inconsistent protection, and doesn't protect all data categories equally. The paper proposes MDV (Model Defense Variational Autoencoder, a machine learning technique that generates synthetic data rather than using real data) to create virtual auxiliary data that effectively addresses all three problems.

This research presents a method to detect deepfakes (AI-generated fake videos or images of faces) by identifying inconsistencies in how image quality degrades between the background and the manipulated face regions. The approach uses a framework that learns to spot these degradation differences through two connected neural networks (deep learning models), one that creates fake images and another that detects them, working together in an adversarial process similar to a GAN (generative adversarial network, where two AI systems compete to improve each other). The method shows better performance when detecting deepfakes created by new, unseen manipulation techniques.

Anthropic patched a vulnerability in Claude Code's network sandbox (a restricted environment that controls where the AI can send data) that could have allowed attackers to bypass security controls and steal sensitive information. The vulnerability, called a SOCKS5 hostname null-byte injection issue (a trick where attackers hide a malicious server address using special characters to fool the security filter), was silently fixed in version 2.1.88 released on March 31, 2025, but was never publicly disclosed or assigned a tracking identifier.

A malicious version of Bitwarden CLI was published on npm for 90 minutes in April 2026, stealing developer credentials through a compromised GitHub Action (an automated workflow tool). The incident received a CVE (common vulnerabilities and exposures, an official vulnerability identifier), but the CVE only notified defenders after the fact rather than providing a patch to apply, highlighting how CVE has drifted from its original purpose of identifying code flaws with fixable versions to tracking security incidents.

NVIDIA Triton Inference Server has a vulnerability in its DALI backend (a component that processes data) that allows attackers to cause uncontrolled resource consumption, potentially leading to a denial of service attack (making the service unavailable to legitimate users).

NVIDIA Triton Inference Server has a vulnerability in its DALI backend (a component that processes data) where an attacker could trigger an integer overflow (a bug where a number exceeds the maximum value a system can store). This could allow an attacker to execute malicious code, modify data, or crash the service.