New paper: “Self-Jailbreaking: Language Models Can Reason Themselves Out of Safety Alignment After Benign Reasoning Training.” Abstract: We discover a novel and surprising phenomenon of unintentional misalignment in reasoning language models (RLMs), which we call self-jailbreaking. Specifically, after benign reasoning training on math or code domains, RLMs will use multipleRead More »Research on Models Engaging in Genie-Like Behavior
The browser update resolves several critical-severity memory safety and memory corruption flaws. The post Chrome 154 Patches 108 Vulnerabilities appeared first on SecurityWeek.
Debates over the plausibility of these doomsday scenarios have heated up since several executives endorsed slowing the technology’s development for safety reasons. The post A Look at AI Doomsday Scenarios That Researchers Say Could Put Humanity at Risk appeared first on SecurityWeek.
Emerging from stealth with $16 million in pre-seed funding, Outerlimit offers a decentralized authorization layer designed to discover, observe, and block harmful autonomous AI actions. The post Outerlimit Raises $16 Million to Stop Rogue AI Agents From Causing Harm appeared first on SecurityWeek.
Remote attackers could trigger the critical-severity flaw to access privileged internal functionality. The post Arista Urges Immediate Patching of Exploited VCO Zero-Day appeared first on SecurityWeek.
Attackers are exploiting a critical flaw in F5 BIG-IP Access Policy Manager (APM) that lets them run code on a BIG-IP system without logging in, F5 says. The flaw, CVE-2026-94127, affects only systems in which APM serves as an OAuth authorization server, issuing access tokens to applications. F5 disclosed it in anRead More »F5 Patches Critical BIG-IP APM Zero-Day Exploited for Unauthenticated RCE on OAuth Servers
A Chinese threat actor codenamed UTA0565 has been observed exploiting the recently disclosed Google Chrome-Microsoft Windows exploit chain as zero-days through fake websites. The attacks, detected on September 3 and 4, 2026, involved the chaining of two vulnerabilities in Chrome (CVE-2026-85046, CVE-2026-87491) and one impacting Windows Advanced Local Procedure CallRead More »Chinese Hackers Exploit Chrome-Windows Zero-Day Chain to Deploy CLEANGULP Malware
Unauthenticated attackers could send malicious traffic to BIG-IP to achieve remote code execution. The post Critical F5 BIG-IP Vulnerability Exploited as Zero-Day appeared first on SecurityWeek.
The cybercrime group is unhappy with its description in an FBI report and threatens to leak stolen information. The post ShinyHunters Claims FBI Hack, Demands Retraction of Threat Report appeared first on SecurityWeek.
A new security vulnerability in Next.js could allow attackers to run code on a server via ImageResponse, the feature that generates Open Graph and other social preview images, Vercel said. The risk applies when an app puts values an attacker controls, such as text read from the request URL, into theRead More »Critical Next.js ImageResponse Flaw Can Lead to Server Code Execution via Crafted SVG Input
