AI Safety Vulnerabilities Escalate: Encrypted Reasoning Cracked, SkillJack, and Systemic Failures

Agent security failures now span encrypted-context privacy, persistent backdoors, memory distortion, hallucination, and workflow compliance. New evaluations and defenses improve measurement and mitigation, but Anthropic's internal failures and the Grok zero-click flaw show that monitoring assumptions remain fragile; recent rogue-agent news claims are unverified.