AI Sec Watch

Anthropic released Claude Opus 4.7, its most powerful generally available model, which improves performance on complex software engineering tasks, image analysis, and instruction-following compared to the previous version. This release follows Anthropic's announcement of Mythos Preview, a more powerful cybersecurity-focused model designed for security-related tasks.

Google is using its Gemini AI model to detect and block malicious ads on its platforms, removing 8.3 billion ads in 2025 as scammers use cloaking techniques (hiding the true destination of a link) and AI-generated content to create deceptive advertising at scale. Gemini analyzes billions of signals like advertiser behavior and campaign patterns to identify harmful ads in real time, including those impersonating legitimate brands to distribute malware, steal cryptocurrency, or redirect users to phishing sites (websites designed to trick users into revealing passwords or personal information). Google reports this approach has reduced incorrect advertiser suspensions by 80% and plans to expand Gemini's use across more ad formats.

OpenAI has expanded access to GPT-5.4-Cyber, a specialized AI model trained specifically for cybersecurity defense work, making it easier for legitimate security professionals to use it. This move follows Anthropic's release of their own cybersecurity model called Mythos.

ATHR is a cybercrime platform that automates vishing attacks (voice phishing, where attackers trick people into revealing passwords over the phone) using AI voice agents and human operators to steal login credentials from services like Google and Microsoft. The platform handles the entire attack chain, from sending fake security alert emails to using AI-driven phone calls that impersonate support staff and extract verification codes. According to researchers, ATHR makes vishing attacks much easier to launch because it requires less technical skill and manual work than traditional attacks.

AI models are becoming increasingly capable at finding vulnerabilities and generating exploits, which lowers the barrier for attackers and compresses the time between vulnerability discovery and widespread attacks. As threat actors weaponize these AI capabilities, enterprise defenders face a critical challenge: they must harden software rapidly and defend systems that haven't yet been patched, because traditional human-speed security processes will not be able to keep pace with machine-speed threats. The source notes that defenders need to strengthen security playbooks, reduce exposure, and incorporate AI into their security programs.

Investigative reporter Ronan Farrow co-authored a 17,000-word article in The New Yorker examining OpenAI CEO Sam Altman's trustworthiness and his track record of misrepresenting facts to people around him. The reporting documents Altman's role in transforming OpenAI from a nonprofit research lab into a nearly trillion-dollar company, as well as the 2023 incident when the board fired him over alleged lying before quickly rehiring him.

Voice biometric systems (technology that identifies people by their voice) are vulnerable to replay attacks (where an attacker plays back a recorded voice to fool the system), but there hasn't been enough realistic training data to build good defenses. This research created RIRplay, a simulated database that realistically mimics how replay attacks actually happen across different acoustic environments, which improved detection performance significantly when tested on real-world voice spoofing challenges.

Researchers developed AdvFor, a black-box attack method (a way to trick an AI system without seeing its internal workings) that can fool image forgery localization models, which are AI systems trained to detect where images have been fake-edited or manipulated. The attack uses reinforcement learning (a technique where an AI learns by trial and error to maximize rewards) to craft minimal changes to images that make forgery detection fail, using only 7 queries per image, and the researchers tested it on multiple real-world models to show it works effectively.

This research proposes a new method for private set operations (PSO, techniques that let organizations securely compare or combine datasets without revealing private information) that reduces the computational burden on client devices. The approach uses secret sharing (splitting data into pieces so no single party can see the whole picture) to allow servers to do most of the work while clients can stay offline, making it practical for large-scale collaborative research across institutions like hospitals.