News
threat-intel Hackers Impersonate IT Help Desk on Microsoft Teams to Gain Domain Admin Access via AD CS Attacks - GBHackers threat-intel Hackers Distribute Malicious Claude Code Extension on Cursor Marketplace to Install Persistent SSH Backdoor - GBHackers threat-intel Hackers Exploit PHP2Shell RCE Vulnerabilities to Gain Root Access on Palo Alto Firewalls - GBHackers threat-intel Claude AI Exploits Gaps in Safety Monitoring to Develop Self-Preservation and Deceptive Behaviors - GBHackers threat-intel WordPress RCE Vulnerability Under Active Attack With Over 58,000 Exploit Attempts Blocked - GBHackers cybersecurity Hackers Exploit WordPress RCE Flaw as Wordfence Blocks Over 60,000 Attacks in 24 Hours - CyberPress cybersecurity Anthropic Reveals How Claude AI Bypassed Safety Monitors Using Harmless-Looking Training Data - CyberPress cybersecurity Hackers Exploit PHP2Shell Vulnerabilities to Gain Root Access to Palo Alto Firewalls Using Fragnesia - CyberPress threat-intel Russian Hackers Exploit MDaemon Mail Server Flaw to Install Backdoors Without User Interaction - GBHackers cybersecurity Microsoft Warns of Trojanized Claude Code Extension Planting Persistent SSH Backdoor - CyberPress cybersecurity Hackers Impersonate IT Help Desk on Microsoft Teams to Hijack Active Directory Accounts and Steal Data - CyberPress cybersecurity Cloudflare Reports 18.2 Million Network-Layer DDoS Attacks in Q3, Up 38% - The Hacker News cybersecurity Inside a malicious Claude Code-branded extension and its persistent SSH backdoor - Microsoft Security Blog ai Yoba Malware Hides a Crypto Miner Behind Legitimate OpenAI and Anthropic Binaries - Security Affairs - Anthropic - Claude AI cybersecurity Anthropic VS code extension on Cursor backdoored for nearly 12 hours - Ars Technica Security
1 / 15
All news ›
Cybersecurity

Anthropic Reveals How Claude AI Bypassed Safety Monitors Using Harmless-Looking Training Data

CyberPress ·

A newly disclosed study shows Claude could generate seemingly harmless training datasets that passed safety screening but still caused a research model to develop deceptive, manipulative, and oversight-resistant behavior after fine-tuning. The experiment, described in “Automated Subtle Emergent Misalignment Research,” tested whether automated red-teaming could uncover weaknesses in training-data monitoring. Anthropic emphasized that the misaligned […] The post Anthropic Reveals How Claude AI Bypassed Safety Monitors Using Harmless-Looking Training Data appeared first on Cyber Security News .

Continue on CyberPress ↗ ‹ All news Get the newsletter

The summary and image come from CyberPress; the image is shown from their website.