Safety

Anthropic's August threat report frames AI misuse as novel when attackers simply repurpose safety tools for extortion and fraud. Here's why your SOC team is already behind.

Anthropic's grateful update on US CAISI and UK AISI red-teaming hides a confession: Claude Opus 4's Constitutional Classifiers harbored a universal jailbreak. Here's the security fallout for anyone running Claude in production.

A take-home interview project was revealed to be a malicious operation, designed to infect candidates' systems with malware. The project, disguised as a legitimate coding assignment, contained hidden directories and scripts that could compromise a user's system.

TP-Link Kasa cameras have been found to leak home GPS data via unauthenticated UDP requests, exposing user locations. Firmware updates are available to patch the vulnerability.

A new wave of AI-powered voice fraud is sweeping the nation, leaving law enforcement and financial institutions struggling to keep up. With just three seconds of audio, scammers can create a convincing clone of a person's voice, allowing them to trick victims into handing over sensitive information or money.

Noma Labs discovered a critical prompt injection vulnerability within GitHub's new Agentic Workflows, allowing an unauthenticated attacker to silently pull data from private repositories. The vulnerability, named GitLost, highlights the security challenges of agentic AI systems.

A use-after-free vulnerability in OpenBSD allows local privilege escalation to root, posing a significant security risk. The vulnerability, identified as CVE-2026-57589, affects OpenBSD versions through 7.9.