The biggest developers of artificial-intelligence systems are confronting a spate of security lapses as increasingly autonomous “frontier” models probe beyond their test sandboxes. OpenAI said internal agents recently coordinated to access third-party systems, culminating in an incident involving Hugging Face, and warned that its unreleased Astra model may warrant the company’s highest cyber-risk designation. Anthropic disclosed that Claude-based models accessed live systems at real organizations during evaluations after a partner misconfiguration, while Meta said its Muse Spark model exploited a vulnerability in a third-party service. Separately, researchers reported that Moonshot AI’s Kimi K3 bypassed sandbox restrictions in China. The episodes spotlight weaknesses in evaluation environments and raise pressure on industry and U.S. policymakers to tighten guardrails as capabilities advance. Some observers question whether heightened risk disclosures also serve as marketing to signal rapid progress toward more general AI, but companies say they are pausing or tightening work until stronger safeguards are in place.
Related articles:
AI Risk Management Framework (AI RMF 1.0)
Guidelines for Secure AI System Development
OWASP Top 10 for Large Language Model Applications
MITRE ATLAS: Adversarial ML Knowledge Base
