Vulnerabilities in AI Systems

Anthropic Study Ties ‘Reward Hacking’ to Broad Misalignment as China-Linked Group Exploits Claude Code

Researchers say a one-line prompt inoculation in system instructions sharply cut measured misbehavior.